Customers

SOCRadar powers rapid threat detection with AlloyDB and Gemini Enterprise

Wed, 01 Jul 2026 19:00:00 +0000

Editor’s note: SOCRadar is a leading cybersecurity company that provides threat intelligence to businesses worldwide. As the volume of cyber threats continued to grow, SOCRadar needed to modernize its data infrastructure to deliver faster insights to its customers. By migrating from PostgreSQL to AlloyDB, SOCRadar achieved a 20x performance boost, reduced operational overhead, and is now better positioned to innovate and grow.

How SOCRadar supercharges rapid threat detection with AlloyDB

SOCRadar provides external threat intelligence to help organizations across 30+ countries defend against cyberattacks. On the front lines of cybersecurity, timely intelligence is everything and a delay of a few minutes can mean the difference between a blocked exploit and a full-scale breach.

As SOCRadar’s business scaled and cyber threat volumes exploded, their on-premises, self-managed PostgreSQL database hit a wall. The database simply couldn't keep pace with the simultaneous demands of high-velocity data ingestion and heavy, real-time analytical queries. This created a severe data bottleneck, slowing down the delivery of critical insights to customers and pulling engineers away from innovation to focus on constant manual database tuning.

Evaluating database alternatives: The hunt for scalability

The engineering team realized their traditional PostgreSQL environment had reached its absolute performance limits. To scale, SOCRadar needed a high-performance fully managed database that could dramatically slash operational overhead while elegantly handling a complex, hybrid workload.

They evaluated alternatives and selected Google Cloud's AlloyDB for PostgreSQL. Because AlloyDB is fully PostgreSQL-compatible, it offered a low-risk migration path while promising a specialized architecture built to handle both high-volume transactions and real-time analytics simultaneously. To accelerate the transition, SOCRadar partnered with NGC, a Premier Business Partner, who meticulously validated the architecture before executing a precision cutover with minimal downtime.

Taming a "triple-threat" workload

Migrating to AlloyDB transformed how SOCRadar processes massive, diverse cyber telemetry. Today, AlloyDB effortlessly manages what SOCRadar’s engineering team calls a "triple-threat" query environment, maintaining sub-second lookup latency even as processing volumes scale.

To understand the performance leaps, it helps to separate the system’s velocity (handling live data streams) from its depth (analyzing historical data):

High-Velocity Transactional Ingestion (OLTP): The platform constantly ingests real-time telemetry from thousands of disparate, fast-moving sources—including Dark Web forums, botnet logs, and social media feeds. AlloyDB handles these continuous INSERT and UPSERT operations with a 3.2x boost in live ingestion velocity, ensuring that the newest threat indicators are immediately recorded and available for detection.
Real-Time Operational Point-Reads: When a security analyst is actively investigating a live incident, speed is everything. Baseline performance testing under zero-load conditions for random ID lookups on indexed fields (e.g., querying a specific Indicator of Compromise by ID) showed that standard queries requiring 3 to 3.5 seconds were completed in just 1 second on AlloyDB.
Deep Analytical Aggregations (OLAP): When a client requests a complex sectoral report such as correlating the most prevalent attack vectors in the finance sector over an entire year, the database must execute deep scans across vast historical datasets. Leveraging AlloyDB’s built-in In-Memory Columnar Engine, these analytical queries run up to 20x faster than standard PostgreSQL.

More than just speed: Reclaiming 45 TB and 75% of DBA time

While the raw performance gains were massive, the operational and financial impact completely changed how SOCRadar's engineering team works day-to-day.

Thanks to AlloyDB's advanced automation, including intelligent memory management and write-ahead log (WAL) optimization, the need for constant, manual database tuning evaporated. The database administrator's (DBA) workload dropped significantly, requiring a system health check just “about once every two or three days." This freed up 75% of SOCRadar’s DBA resources, allowing them to pivot away from maintenance and focus entirely on core platform innovation.

Financially, AlloyDB’s dynamic storage management solved a massive cost efficiency issue. Unlike traditional database environments that lock you into paying for fixed, provisioned storage even after data is purged, AlloyDB automatically scales storage down to match actual data footprints. By clearing out legacy, unnecessary logs, SOCRadar was able to instantly reclaim over 45 TB of storage, achieving massive, automated cost optimization.

Fighting alert fatigue with integrated Gemini Enterprise Agent Platform

Beyond scaling infrastructure, AlloyDB has allowed SOCRadar to redefine the core architecture of their threat response using artificial intelligence.

Security operations centers (SOCs) globally are plagued by "alert fatigue"—the sheer volume of security alarms makes it easy to miss a critical attack. To solve this, SOCRadar integrated Gemini Enterprise Agent Platform as a core component of their solution architecture, linking it directly to their Alarm Management framework running on AlloyDB.

By running Gemini AI-native filtering directly on their active data workloads, SOCRadar can automatically distinguish between true positives and benign false alarms. The AI categorizes, filters, and routes alerts before they ever reach the end-user. This ensures security analysts are insulated from noise and receive only the most critical, validated, and actionable intelligence.

Expanding capabilities: The future of agentic threat hunting

With a high-performance foundation firmly established, SOCRadar’s dedicated AI team is transitioning from passive analytics to active automation. The company is currently testing Agentic AI workloads, with plans to roll them into production in subsequent phases.

By integrating Real-time Data Agents with Gemini Enterprise and AlloyDB, SOCRadar is transforming with autonomous agents that don't just store data, but actively hunt threats, reason over context, and take action. Their upcoming production roadmap includes:

Natural Language Querying (NLQ): Allowing analysts to conduct rapid threat hunting using conversational language, lowering the technical barrier to querying massive database sets.
Intelligent Semantic Similarity Search: Leveraging native vector embeddings and Gemini Enterprise to allow Data Agents to independently surface hidden patterns across historical logs that traditional keyword searches would miss.
Automated Incident Summarization: Instantly transforming hundreds of lines of complex, deeply technical logs into concise, plain-language executive summaries for security analysts during critical incidents.

By consolidating transactional velocity, historical depth, and built-in AI intelligence into a unified platform, SOCRadar has eliminated its data bottlenecks and built a highly automated, future-proof framework for global cybersecurity defense.

Ready to modernize your database infrastructure? AlloyDB provides a fully managed, PostgreSQL-compatible database with high performance for transactional, analytical, and AI workloads. Learn how you can reduce costs, eliminate management overhead, and build intelligent applications.

How Schrödinger sped up molecular discovery by 4x with Alphaevolve

Tue, 30 Jun 2026 16:00:00 +0000

Computational chemistry researchers have traditionally faced a frustrating trade-off when simulating molecular interactions: use fast classical force fields that sacrifice precision or rely on accurate quantum-mechanical methods that run too slowly on large jobs.

Machine-learned force fields (MLFFs) close that gap by training neural networks on high-fidelity quantum data. When it comes to modern drug discovery and materials design, though, there’s demand for even faster processing speeds to handle massive chemical libraries involved. To overcome such performance constraints, Schrödinger partnered with Google Cloud to deploy AlphaEvolve, an evolutionary AI coding agent developed by Google DeepMind that iteratively generates and refines algorithms to find the most efficient code path overcoming the algorithmic bottleneck.

A collaborative duet with AlphaEvolve

Schrödinger — a leader in developing scientific software for over three decades — identified two critical algorithms within their MLFF training pipeline that limited performance: neighbor list computation and Ewald summation. These algorithms aggregate data from atomic neighbors and calculate long-range potentials, but both became limiting factors in training and inference speed.

Schrödinger's primary technical goal was speeding up AI model training for energy and force calculations. Specifically, they targeted the Ewald summation, a critical but computationally demanding function used in molecular mechanics. The Ewald sum was the main performance constraint in Schrödinger's PyTorch code. It had no established vectorized algorithm and often relied on simple for-loops that ran slowly on large simulations.

By incorporating AlphaEvolve into their models, the system could generate a batched implementation of the Ewald summation using parallel batch matrix multiplication. This would evolve the PyTorch code to outperform existing custom kernels.

Evaluation metrics

Schrödinger used a rigorous multi-layered evaluation framework to confirm the evolved code was both performant and scientifically accurate:

Inverse time (primary metric): The core objective was to maximize throughput by reducing calculation time, from a baseline score of 7.9.
Functional correctness: All evolved programs had to pass a full test suite, including regression tests on complex systems such as disordered water models.
Success rate: This was measured by the share of programs that were both functionally correct and faster than the baseline.

“AlphaEvolve allows us to explore larger chemical spaces faster and more efficiently than ever before. Faster MLFF inference carries real business impact, shortening R&D cycles in drug discovery, catalyst design, and materials development, and enabling companies to screen molecular candidates in days rather than months.” — Gabriel Marques, technical lead of machine learning, Schrödinger

Results: a 4x speedup and breaking bottlenecks

By applying AlphaEvolve, Schrödinger replaced simple for-loops in the Ewald summation code with parallel batch matrix multiplication. This optimization raised the program success rate from less than 1% (40 out of 5,000 evaluations) to more than 60%, while improving the performance metric from the baseline of 7.9 to nearly 30.

Optimizing these foundational algorithms delivered a 4x speedup in both MLFF training and inference. This acceleration lets researchers compress molecular screening timelines and directly benefits several key research areas:

Drug discovery: Identifying viable therapeutic candidates quickly to address urgent medical needs.
Catalyst design: Developing efficient chemical processes for industrial applications.
Materials development: Designing next-generation materials with custom properties for electronics and energy storage.

The next evolution

Schrödinger plans to apply this evolutionary approach to custom GPU kernels to test whether AI-generated code can outperform human-engineered implementations.

Read the full technical paper on AlphaEvolve to learn how evolutionary AI agents optimize scientific codebases, or contact the Google Cloud AI team to discuss accelerating your research workflows.

How growing UK midsize businesses are building in the AI era

Wed, 17 Jun 2026 08:00:00 +0000

The UK’s 5-million-plus small and midsize businesses and enterprises (SMBs) are the backbone of our economy. Today, we’re seeing these critical businesses begin to put AI to work, to operate more efficiently, move faster, and ultimately deliver better outcomes for their customers.

This shift is driven by tangible day-to-day results. According to recent research from Enterprise Nation published in partnership with Google, 71% of AI adopters surveyed in the UK say the technology helps them save time on routine tasks, and 64% report a direct boost in productivity. On top of this, AI-enabled productivity tools (like Google Workspace with Gemini) are delivering a 20% boost in productivity for SMBs, which effectively hands them back one full working day every single week.

At Google Cloud, we have a front row seat to this shift: SMBs have long utilized platforms like Google Workspace, and today they’re transforming with Google’s AI platform and models. In fact, we’ve seen the number of UK-based SMBs using Google Cloud AI nearly double year-over-year. This includes our Gemini models and products like Gemini Enterprise and AI Studio, which are helping SMBs do things like:

Roll out better customer support systems to help escalate and resolve customer support calls more quickly.
Automate repetitive actions in areas like payroll and accounting.
Help more employees understand and leverage data at work — even those not trained as data analysts.
Rapidly create and implement new designs for marketing collateral.
Help more people build their own AI agents to help them in their everyday jobs.
Conduct complex research projects at a speed and price point previously unavailable.

At today’s Google Cloud London Summit, we’re showcasing a number of innovative SMB customers who are actively using our AI tools to transform how they work, including companies who have recently expanded their work with us:

Neural Alpha, a sustainability fintech company, is using Gemini models to read unstructured environmental and corporate sustainability reports to automatically find and organize thousands of key facts, cutting months of slow, manual research down to a fraction of the time.
Sep 2, a digital security provider, uses Gemini Enterprise to deploy autonomous AI agents for 24/7 threat monitoring — accelerating incident detection and quickly neutralizing security threats for its customers.
Sunhouse, a strategic brand design agency, uses Gemini Enterprise to easily find archived design work stored on Google Drive, enabling its teams to spend less time hunting for files and more time growing its business with global brands.
Terrapinn, a global B2B events company, is transforming its operations by leveraging Gemini models, NotebookLM, Looker, and BigQuery to turn manual tasks into automated workflows, accelerating how its teams design, market, and deliver world-class conferences.
VoCoVo, a telecommunications provider, is integrating Google Cloud AI across its systems to turn isolated data into actionable intelligence and build autonomous workflows, streamlining routine operations so their team can focus on high-impact innovation.

Empowering Your Team: AI Upskilling Resources for Growing British Businesses

To help midsize teams maximize their impact and confidently navigate the modern AI landscape, we’ve developed a suite of dedicated, no-cost upskilling resources. Whether you want to train your existing teams or democratize data tools across your entire workforce, these programs will help you build an AI-ready organization:

SMB-Focused Programs: Explore our new SMB Learning Path or enroll in the Gemini Enterprise Agent Ready (GEAR) program for specialized training in agentic AI.
Google Skills for Organizations: Access our no-cost, on-demand learning platform featuring over 3,000 AI courses and hands-on labs created by experts at Google Cloud and Google DeepMind.
Get Certified: Ready to validate your team's expertise? This premium, cohort-based program offers instructor-led training, technical mentorship, and AI-infused skill badges designed to prepare your team for industry-recognized certifications.

By offering a full suite of SMB technology and training — from productivity in Workspace, to all our Ads services, and now powerful AI tools — Google is helping small and midsize firms thrive, no matter where the future takes us.

From AI potential to agentic reality: Driving the UK’s next chapter

Wed, 17 Jun 2026 08:00:00 +0000

The United Kingdom, and London in particular, continues to be one of the great hubs for AI development in Europe and the world. We’re home to Google DeepMind, of course, as well as significant AI unicorns — and Google Cloud customers — like Ineffable Intelligence, which is today announcing an important partnership with us.

A year ago, we joined you for the London Summit to showcase the vast potential of generative AI, including a major investment in upskilling the UK civil service. Today, as we welcome our partners once again to the historic vaults of Tobacco Dock, that potential has become an industrial-scale reality. In my conversations with leaders across both Whitehall and The City, the focus has moved from chatbots and media experiments to full-production execution. This is the moment of the agentic enterprise, where we shift from systems that simply chat with us to systems that can reason, plan, and execute multi-step workflows.

This transition is the cornerstone of the UK’s projected £400 billion economic boost from AI by 2030. At Google Cloud, we are the only provider offering the full integrated stack — custom silicon, frontier models, and planet-scale infrastructure — required to turn the Agentic Enterprise into a reality.

The new frontier of British enterprise and research

The banking sector is a key proving ground for this shift. And HSBC, one of the largest and most important financial institutions in the world, is showing the way. Today, we’re announcing a multi-year transformational partnership with HSBC to accelerate AI adoption across HSBC’s products and services globally. This new collaboration will further accelerate the shift towards AI-enabled ways of working across HSBC’s global operations.

HSBC will work with Google Cloud and Google DeepMind engineering teams to collaborate on new AI-powered tools and programmes, with access to Google’s latest agentic AI capabilities – including Gemini models and the Gemini Enterprise Agent Platform. The initial delivery focus on three areas: hyper‑personalised wealth management support, stronger financial crime risk management, and AI tools to enhance frontline/relationship manager client service

UK startups also continue to break new ground with technology, and AI in particular, as demonstrated by the work of frontier labs like Ineffable Intelligence. The company, which launched earlier this year, has chosen Google Cloud as its preferred cloud partner, utilizing Google’s full stack of AI-optimized hardware and tools to build and train Ineffable’s first generation of foundational models.

Led by David Silver, a former Google DeepMind researcher who was instrumental in the AlphaGo project, Ineffable Intelligence is taking a unique approach to AI development. The team are building systems that learn primarily through their own experience through reinforcement learning, instead of relying on the large-scale human-generated datasets behind language models. The ambition is to create a “superlearner” that develops knowledge through trial and error. This year, Ineffable Intelligence set a record for a European seed funding round of $1.1 billion, and now Ineffable Intelligence will support its training work by deploying one of the largest clusters of A5X, powered by the NVIDIA Vera Rubin NVL72 platform on Google Cloud, delivering massive computational scale.

To move from experimentation to true industrial production, businesses need more than just models; they need a roadmap. To help show them the way, we’re expanding our partnership with Deloitte, which will open a new AI Studio at its London campus. Developed in collaboration with Google Cloud, the studio will help British organisations move beyond AI experimentation to deploy autonomous, action-oriented AI systems at scale.

Deloitte is also committing to upskill 1,000 members of its UK AI and data workforce on Gemini Enterprise. This certification program will ensure that Deloitte’s AI and data engineers’ are equipped with the technical expertise to implement Google’s most advanced agentic architecture, providing UK clients with one of the largest pools of certified AI talent in the region.

Building a future-ready public sector

The blueprint for a modern digital government requires moving away from rigid legacy contracts toward agile, AI-driven public services. In collaboration with the Ministry of Housing, Communities and Local Government (MHCLG), the i.AI incubator, Google Deepmind, and Faculty, we are delivering tangible public sector reform and tools for reinvention that directly support the national goal to "get Britain building."

Agencies like MHCLG are already using a tool called Extract which was built using Google technology to help transform planning processes by reducing document processing times from two hours to just two minutes. Simultaneously, we are supporting trials of an AI planning tool — co-created with local planning authorities in Barnet, Dorset, and Camden — which aims to cut decision times for everyday applications by 50%. Furthermore, the Department for Transport (DfT) is utilizing Gemini to streamline public consultation analysis, a move projected to save £4 million annually.

Innovation on this scale also requires a secure, sovereign foundation. That is why Google Cloud is working to strengthen our UK data residency commitments, including measures like making Gemini 3.5 Flash, which features in-country AI processing, available by late June 2026 for sensitive sovereign use cases. We are giving British organizations the confidence to innovate within strict compliance boundaries.

To help keep businesses safe from the challenges posed by bad actors using AI and other digital threats, we also recently announced a comprehensive AI-powered cybersecurity platform — Google AI Threat Defense — which combines Wiz, Mandiant, Gemini & CodeMender to find, fix, and protect our customers from vulnerabilities.

Proven impact from the high street to public service

Autonomous agents are no longer a future prospect; they are delivering value across the UK economy today. Our work with THG Ingenuity, an ecommerce solutions provider, has delivered an 8x higher conversion rate via its AI Shopping Assistant. Starling is similarly empowering customers with "spending intelligence" tools for instant habit analysis around purchases and expenses. And Rightmove, has launched a beta version of an AI-powered conversational property search, built with Google’s Gemini models, enabling users to search for homes in their own words.

The breadth of this impact is visible across every sector: Kingfisher is pioneering retail-specific agentic applications; Openreach is driving field service optimization in telecommunications; andUnilever is using AI at scale across the entire value chain to drive growth and build desirable brands in the new era of consumer goods.

Meanwhile, VMO2 is streamlining complex data operations; Vodafone is executing a $1 billion partnership to redefine network performance; and WPP is integrating Gemini across creative workflows, whether that's generating high-fidelity campaign assets at speed and scale, powering AI agents, or training robotic camera operators.

Empowering the engine of growth for small to medium businesses and startups

The true measure of Britain’s AI success lies in its small and medium enterprises and startup ecosystem. Our AI Works research highlights a pivotal moment: AI has the potential to boost productivity for small and medium enterprises by 20% and unlock £198 billion in output for the UK economy. With 56% of smaller firms already seeking guidance, we have launched the AI Works for Britain upskilling initiative to ensure no business is left behind.

We also continue to foster the next generation of British unicorn startups through our ongoing partnership with Tech Nation at the London AI Hub. This sustained commitment ensures founders have the resources and community needed to scale, and this September, we will further this mission by hosting the Gemini Startup Forum: Cybersecurity in London to help startups build secure-by-design AI applications.

The Model Garden at Platform 37

Our belief in the UK’s potential is reflected in our physical footprint, too. We are continuing to invest in the UK's digital infrastructure to support growing demand: Our state-of-the-art data center in Waltham Cross launched in September 2025, a key part of our two-year, £5 billion investment to help power the UK's AI economy. And earlier this year, we opened our new office in London in Kings Cross, Platform 37, along with plans for The AI Exchange, a new public space dedicated to deepening understanding of AI.

Building on this momentum, we are excited to introduce The Model Garden at Platform 37, launching in the fourth quarter of 2026. This London-based hub is far more than a physical space; it serves as a strategic investment designed to fundamentally elevate how we engage with our most important customers. Blending the timeless aesthetics of a classic English garden with immersive, high-tech innovation — from living digital walls to a three-story atrium — The Model Garden acts as a physical marketplace for our best ideas.

The blueprint for the agentic enterprise

For UK businesses, civic leaders, and organizations to continue to lead in the AI moment, they must not only rethink the technology they use but also fundamental aspects of how we work. As we support thousands of organizations and millions of teams here and around the globe, we see three core strategies helping achieve success with AI:

Culture: We must reimagine our organizations for the future. True transformation means getting teams excited, enabled, and equipped to work with AI agents in completely new ways. It is about human-AI collaboration, not just automation.
Responsibility: We must build with safety and security in mind from day one. Protecting your users, your customers, and your brand is paramount. Our frontier models are built on a foundation of rigorous AI principles and secure-by-design infrastructure.
Sustainability: In an era of rising compute demands, we must scale in a way that is both financially viable and positive for our planet. At Google, we are committed to carbon-free energy 24/7, ensuring that the UK’s AI growth does not come at the cost of our climate goals.

Architecting the future together

Google Cloud is the primary partner for the UK’s agentic transition. We are moving beyond the hype of experimentation into the rigor of production. From the research labs of King's Cross to the diverse enterprises powering the high street, we are architecting a resilient, sovereign, and prosperous future for the United Kingdom.

Thank you to everyone who’s joining us in London — yesterday, today, and into the future. This year we’ve packaged up an exclusive on-demand experience, allowing you to stream the defining London Summit moments, available anywhere, anytime.

How Atlas scales hundreds of merchant databases with Cloud SQL Enterprise Plus edition

Tue, 16 Jun 2026 16:00:00 +0000

Atlas is building the operating system for restaurants. Online storefronts, point of sale, third-party logistics, food platform integrations, customer loyalty, and AI tools represent everything a restaurant needs to start, run, and grow. We work with brands like SaladStop, Killiney, Haidilao, Raffles Hotel, Lo and Behold Group and the Les Amis Group in Singapore, helping merchants increase basket sizes, grow sales, and reduce operational costs.

Every merchant on Atlas gets their own dedicated Cloud SQL for PostgreSQL database. Restaurants are very different from each other. A single-outlet cafe and a multi-outlet chain should not look the same underneath. Isolated databases give us full data separation, predictable performance even during peak lunch and dinner rushes, and the flexibility to scale, tune, or migrate each merchant independently. As Atlas grows, the number of databases grows with us.

The challenge: Scaling beyond standard

We started on the standard Cloud SQL Enterprise edition. It was a solid foundation, but as we onboarded more merchants and shipped more features, the operational layer required to manage our databases became a bottleneck.

We were managing connection pooling as a separate layer, which meant more services to run, secure, and monitor. When a query caused a CPU spike, we needed to know exactly what happened and which merchant triggered it, but we were spending too much time reconstructing problems from limited signals. With a lean team and no dedicated database engineers, every extra component multiplied the maintenance load.

The shift to Enterprise Plus edition

When we needed to provision new database instances, the Google Cloud team introduced us to Cloud SQL Enterprise Plus edition. We were already asking ourselves how much more operational overhead this was going to add, and what stood out was that Enterprise Plus edition removed whole categories of work we would otherwise have to own.

Managed connection pooling: Now built directly into Cloud SQL, we no longer run pooling as a separate layer. This means fewer moving parts, less to maintain, and a smaller security surface area.
Query insights: This was the most impactful feature for our needs. We can now see exactly which queries are expensive and which merchant is triggering them. It turns performance tuning from guesswork into something concrete and actionable. For a platform running hundreds of databases, this visibility is a "superpower."
Data cache: This keeps read performance consistent even as merchant datasets grow. Since restaurants generate more data every day, the data layer needs to stay fast as that complexity compounds.
Near-zero downtime scaling: We can now scale instances as merchants grow without disrupting service during off-peak hours.

After seeing the results on the new instance, we migrated all our existing databases to Enterprise Plus edition as well.

The impact: Focus on innovation, not plumbing

Atlas today powers thousands of restaurant outlets, processes tens of thousands orders daily using hundreds of managed databases. The biggest change is where engineering time goes. We spend 30% less time on database operations and more time building products. Merchant onboarding got simpler because a new merchant is provisioned in seconds with a ready-to-use managed database. We are much more proactive on performance now, catching and fixing issues before they reach merchants. Day to day, we are not thinking about database plumbing. We are thinking about how to serve merchants better and that has allowed Atlas to grow 200% to 300% year over year.

Looking ahead: An AI-first future

We are investing deeply in AI, both internally and externally. Internally, we have gone all in on agentic engineering through AI-assisted development workflows that let a lean team build, review, and ship code significantly faster. Externally, we are building AI-powered tools that help restaurant operators make better decisions and act on them. We have a lot of experimental ideas on the roadmap, including new product surfaces and new ways to help restaurants grow. The thing that gives us confidence to move fast on all of this is that the foundational layer, Cloud SQL and Google Kubernetes Engine (GKE), is battle-tested and does not get in the way.

Google Cloud handles the infrastructure complexity. Atlas stays focused on building the best tools for restaurants.

Cloud SQL Enterprise Plus gave us a database architecture that is flexible, observable, and easy to scale. We are not thinking about infrastructure anymore, we are thinking about our merchants. As we go deeper on AI and continue growing the platform, Google Cloud gives us the confidence to move fast without worrying about what is underneath.

Ready to scale your database architecture?

Don't let infrastructure bottlenecks slow down your innovation. Whether you are managing tens or hundreds of databases, see how Google Cloud SQL can streamline your operations, enhance observability, and give your engineering team the freedom to focus on what matters most.

Explore Cloud SQL Enterprise Plus edition today
Sign up to try Cloud SQL for free.

How Siemens "slices the elephant," advancing agentic workflows for industrial software development

Tue, 16 Jun 2026 07:00:00 +0000

For technology companies like Siemens, software is the nervous system of factories, energy grids, and transportation networks worldwide.

As a global leader in industrial AI, industrial software, and industrial automation, Siemens brings decades of domain expertise across factory and process automation, energy infrastructure, and intelligent transportation — expertise that no off-the-shelf AI solution can replicate. But innovation carries a heavy anchor: legacy code.

With codebases spanning hundreds of millions of lines developed for over more than a decade, Siemens faced a challenge that standard AI tools couldn't solve: understanding and modernizing this code and the applications which run on it. The scale and depth of industrial-grade software demand a fundamentally different approach. Existing coding assistants lacked the contextual depth required to navigate complex, multi-layered industrial codebases — a gap Siemens set out to close.

To solve this, Siemens and Google Cloud created Knowledge Fabric, an AI system for automating the software development lifecycle. It was built using knowledge graphs on Spanner Graph, the Google Agent Development Kit, Gemini API, Gemini Enterprise Agent Platform, Gemini CLI, and Anthropic Claude Code. In a pilot migrating existing frontiers to web-based interfaces, Knowledge Fabric reduced implementation effort, freeing engineers to focus on customer innovations while maintaining full system compatibility.

“By ingesting the entire software ecosystem into an intelligent agentic system equipped with custom knowledge graphs, we aren’t just helping developers optimize their development time; we are enabling autonomous agents to reason across the past to build the future,” said Franz Menzl, senior vice president, product creation excellence at Siemens. “This is about freeing engineers from repetitive work so they can focus on higher-value problem solving.”

The challenge: the complexity of industrial software

Modernizing large-scale industrial-grade software systems is often compared to rebuilding a jet while flying it. For Siemens, the challenge had four dimensions:

Scale: The repositories are massive — far exceeding the context windows of standard large language models.
Fragmentation: Critical knowledge was scattered across code, Jira tickets, Confluence pages, and scanned PDF manuals from the early 2000s.
Complexity: Tracing the link between a specific line of code and a functional requirement document from 10 years ago presented a challenge that no manual or conventional tooling approach could address efficiently. It’s a reality shared across the industry.
Responsibility: Systems must adhere to strict quality, compliance, and lifecycle requirements, often over 15 to 20 years of operation. AI‑generated outputs must therefore be explainable, traceable, and verifiable. Hallucinated or unvalidated changes are not merely inefficient but operationally unacceptable.

"We realized that standard RAG (retrieval-augmented generation) wasn't enough," said Agata Gołębiowska, technical lead, Google Cloud. "Code isn't just text; it has inherent structure. A class belongs to a file, which belongs to a module. Flattening that into a vector database meant losing the representation of relationships elements of the codebase."

The solution: A domain-aware Knowledge Fabric

To make this sprawling software environment navigable for AI-driven workflows, the teams built the Knowledge Fabric agent. This agent goes beyond keyword matching to “understand” the relationships between assets.

We use Spanner Graph to model the inherent structure of the codebase, applying the same rigor to documentation across formats. By mapping connections between these domains, we can link specific code snippets directly to requirements in a design document. Agents then traverse this graph, using tools to query the structure via Graph Query Language (GQL).

But GQL is only one piece. To enable semantic understanding, we generate embeddings for every node, using Spanner's Approximate Nearest Neighbors (ANN) algorithm to perform efficient vector search across the full codebase. Finally, we give agents full-text search capabilities, which can be combined with GQL to pinpoint nodes and edges with precision.

Combining these three methods lets an LLM agent answer complex queries, such as: "Which functions need to be updated if I change the logic in the Axis Control Panel?" The system traverses the graph — weighing keyword and semantic similarity — to identify dependencies, retrieve relevant documentation, and present a precise impact analysis.

This precise context is what lets a coding agent produce a valid, usable, and maintainable implementation.

"Slicing the elephant:" the agentic workflow

A key insight from the project was that AI agents struggle with massive, ambiguous tasks. To succeed, the team adopted a design pattern dubbed "slicing the elephant."

The system breaks a sweeping request like “refactor this module” into smaller, more manageable tasks, each handled by a specialized agent built with the Google Agent Development Kit (ADK):

Search agent: Acts as a deep-research specialist. It uses tools to explore the code graph and cross-reference findings with documentation in Agent Search.
User story agent: Interviews the product owner to gather requirements, then drafts detailed user stories with acceptance criteria linked to existing system contexts.
Architecture impact agent: Analyzes proposed changes against the graph to predict side effects before a single line of code is written.
Task breakdown agent: Consumes the analysis from the architecture impact agent and breaks the work into small, manageable tasks, each carrying all the context relevant to a specific change.
Coding agent: Implements the change described in a specific task. Reaching this step without context and prior analysis produces unusable code.

The system keeps a human in the loop at every step, which ensures reliable, production‑grade outcomes and keeps engineers focused on meaningful work rather than routine implementation.

"By slicing the elephant — breaking complex refactoring jobs into smaller, agent-led tasks — we observed a significant productivity increase," said Alexander Lomakin, project lead at Siemens. "We essentially gave the AI the roadmap it needed to navigate the complexity."

Pilot results: Faster, more efficient engineering

Developers saw results almost immediately.

Analyzing dependencies for a new feature once required senior engineers to spend several days navigating codebases and legacy documentation. With the Knowledge Fabric, the same work now takes far less time.

In a recent production pilot migrating legacy control panels to modern web‑based interfaces, the Knowledge Fabric reduced overall coding effort while preserving system integrity and industrial quality standards.

Engineers now spend more time creating customer value and less on repetitive work.

Get started

The Knowledge Fabric shows that generative AI can do more than write boilerplate code, it can also help teams modernize the legacy systems their businesses depend on most.

To learn more about building graph-based agents for your own legacy modernization:

Read about Spanner Graph.
Explore Agent Platform and find pre-built production-grade agents on Agent Garden
Check out the Agent Development Kit.
Read more on how Siemens is advancing industrial AI.

Architecting a trusted agentic platform with graph technologies: A Yahoo case study

Mon, 15 Jun 2026 16:00:00 +0000

As enterprises adopt agentic AI, they need to shift from reactive systems of intelligence to proactive systems of action to equip the agents they’re building with the context and performance they need, plus regulator-grade accountability, where every decision is explainable and auditable.

At Google Cloud Next ‘26, we discuss how our Agentic Data Cloud enables a system of action, and Yahoo’s digital media buying platform is a compelling example of this vision. Yahoo partnered with Google Cloud to build its Seller Agent digital media buying platform using Google Data Cloud graph technologies. Seller Agent condenses multi-week manual processes into fully governed, live campaigns that can be executed in just seconds. Ultimately, this agentic platform serves as a powerful blueprint for multiple industries, demonstrating that autonomous systems can operate at remarkable speed while remaining strictly accountable.

"Yahoo's mission is to be a trusted guide through the digital world. In partnership with Google Cloud, we're extending that promise to advertisers: agentic media buying that's fast, transparent, effective, and built to be trusted." - Gabriel DeWitt, Head of Monetization, Yahoo

In this blog, we explore the shift toward agentic AI, examine how Yahoo’s Seller Agent architecture solves for speed and trust in media buying, and show you how to apply this graph-based pattern to build trusted systems of action in your own organization.

Case study: agentic media buying

For years, complex, high-value workflows—like premium digital advertising campaigns—have required weeks of human handoffs, fragmented spreadsheets, and manual analysis. Yahoo recognized that agentic AI could collapse this timeline, allowing agents to plan and execute campaigns in mere seconds. This leap from manual to autonomous execution represents a massive opportunity to reclaim operational efficiency and ensure more of every dollar reaches measurable outcomes.

But simply dropping LLMs into a high-stakes workflow does not solve the problem; an agent attempting to negotiate contracts or ad placements without a deterministic understanding of real-time inventory, pricing rules, and business constraints is prone to hallucinate — potentially resulting in disastrous deals. A trusted agentic platform requires a definitive, real-time source of truth, ensuring it acts on hard facts rather than statistical guesses.

Furthermore, speed and factual grounding are only half the equation. The moment an AI agent starts moving real budgets, it faces scrutiny from regulators who demand instant answers to why specific decisions were made or which policies were applied. Digging through raw system logs after the fact is the wrong control surface for autonomous execution. Real-world systems of action require regulator-grade governance and auditability built directly into the workflow, not bolted on as an afterthought.

The architecture of a trusted system of action

Yahoo's mission has always been to be a trusted guide through the digital world. Agentic media buying extends that promise to advertisers, agencies, publishers, and regulators who entrust Yahoo with their budgets — and expect real accountability. The issue was automating campaign execution in a way that was explainable, governable, and auditable.

To meet this challenge, Yahoo built its Seller Agent as a multi-agent system running on Google Cloud. Buyer requests enter through a planning supervisor agent running on Google Kubernetes Engine (GKE) and orchestrated with Google's Agent Development Kit (ADK). The supervisor decomposes each request into specialized tasks including inventory discovery, audience matching, forecasting, pricing analysis, package recommendation, governance review, and execution. Agents coordinate through the open Agent2Agent (A2A) protocol, while Gemini Enterprise Agent Platform hosts models for embeddings, forecasting, and graph learnings.

But the true breakthrough — what makes autonomous execution both fast and fully transparent — is the platform’s dual-graph foundation. The platform is anchored by two specialized graph systems with an intentional separation of duties: a knowledge graph that’s optimized for acting, and a second context graph for remembering and learning.

"As the industry moves from systems of intelligence to systems of action, the constraint on autonomous AI shifts from model capability to whether a business can trust what an agent does unsupervised. Autonomous systems must record why decisions were made and learn from outcomes. That trust is earned through robust data infrastructure. We built that foundation with Google Data Cloud: a knowledge graph for operational truth in Spanner Graph, a context graph for decision lineage in BigQuery Graph — the blueprint for enterprise-scale agentic platforms." - Swapnil Patel, Senior Director and Head of Monetization Engineering, Yahoo

The knowledge graph: Grounding agents in business reality

Powered by Spanner Graph, Yahoo’s knowledge graph represents its monetization business as a connected operational model, grounding every agent decision in business reality. It models advertising products, placements, audience segments, inventory, contracts, and governance controls as first-class entities and relationships. Crucially, policies live directly within the graph as versioned relationships rather than being buried in application logic. This design allows the system to evaluate products, contractual obligations, consent requirements, and regulatory constraints together in a single, unified graph traversal.

The graph acts as a semantic contract across the agentic platform. During campaign evaluation, an agent can navigate from initial buyer requirements to eligible audiences and governing policies within a single query plan. Gemini Enterprise Agent Platform embeddings enrich these entities with semantic similarity, while graph neural networks contribute inferred relationships. Ultimately, this allows agents to do more than just retrieve available inventory — they understand exactly why it is relevant and help ensure it satisfies all governing constraints.

Yahoo’s knowledge graph ontology, aligned with industry standards like IAB AdCOM

The context graph: creating an auditable memory

Execution at agent-scale is only safe if it is entirely transparent — which is the core function of the context graph. Every time the Seller Agent takes an action, that exact operational span is captured by the BigQuery Agent Analytics plugin. In addition to logging the raw events, the system shapes this evidence into a typed, queryable context graph using BigQuery Agent Analytics SDK utilizing Yahoo's decision-trace ontology, stored in BigQuery Graph.

Consequently, every decision point, candidate package, policy evaluation, specialist-agent delegation, and execution outcome becomes a connected graph of evidence. Because this trace is structured as a typed graph, explaining the agent’s decision making process becomes a simple query. An auditor can instantly trace a decision from the originating campaign brief through every score that’s assigned and policy that’s applied. This transforms autonomous behavior from an opaque process into a fully transparent and continuously improving record of decision-making, helping to ensure absolute accountability.

Yahoo’s context graph ontology

From human to agent scale

For a concrete example of the architecture in action, consider an ad campaign run. What traditionally required weeks of coordination across planning, sales, operations, and compliance can now be completed in seconds through two simultaneous processes.

Acting via the knowledge graph. This pipeline moves the budget, navigating linearly from the buyer's request to a live campaign ground on the knowledge graph. This proceeds in four steps:

Submitting the brief: A buyer agent submits a campaign brief over Ad Context Protocol (AdCP) that describes the desired audience, budget, geography, and business objective.
Knowledge retrieval: The Seller Agent queries the knowledge graph to identify relevant inventory, audiences, contractual availability, historical performance, and governing policies.
Evaluation and scoring: The agent evaluates these factors together to assemble a package of media buying candidates. Forecasting models score the opportunities, while a governance agent independently reviews consent, brand safety, and regulatory constraints.
Approval and execution: The package is either approved automatically under policy thresholds or escalated for human review. Once approved, the media buy is executed and activated.

Auditing and learning via the context graph. While the execution pipeline moves forward, this parallel loop continuously captures the system's reasoning in the context graph, helping to ensure transparency and improve future cycles. This offers the following capabilities:

Continuous capture: Every candidate considered, score assigned, policy applied, and governance decision becomes a connected record in the context graph, linked to the originating campaign session.
Closed-loop learning: As delivery, attribution, and outcome signals arrive, they are joined back to the decisions that produced them, creating the training data that improves future recommendations.
Instant explainability: If an advertiser asks why a particular package was selected or which policies influenced the outcome, the answer is preserved in the context graph and reachable through a single query.

The result is a platform where knowledge, decision-making, governance, measurement, and learning operate together — allowing autonomous media buying to remain explainable, auditable, and continuously improving.

A blueprint for many industries

The era of AI as a mere advisor is ending. Enterprises are demanding systems of action — autonomous agents capable of executing complex, multi-step workflows. But in regulated sectors, the speed that AI brings to the table turns into a liability if you cannot prove how a decision was made. The primary barrier to autonomous execution is no longer intelligence; it is trust.

The architecture that Yahoo and Google Cloud built provides a broadly applicable blueprint with which to solve this. While designed to fix the bottlenecks of digital media buying, the underlying pattern applies to any industry managing high-stakes decisions — from financial trading to supply chain logistics. To operate at agent speed but still maintain human oversight, enterprises must adopt a new architectural baseline that:

Grounds decisions in business reality: Agents cannot rely on probabilistic models alone. They must be grounded by a knowledge graph that deterministically maps your business logic, active contracts, and compliance rules.
Builds an auditable memory: You cannot govern what you cannot trace. Every agentic action must be captured in a context graph, creating an immutable, queryable record of exactly why a decision was made and which alternatives were rejected.
Embraces open interoperability: Trust requires transparency. By building on open protocols and provenance standards, industries can establish a common, auditable language for agentic behavior.

As foundational models become commoditized, enterprises’ competitive advantages are shifting. Long term, your moat will not be the language model you deploy, but the proprietary graph of your business operations and governed history. Likewise, the future of enterprise AI isn’t simply systems that can act, but systems that can explain, govern, and take accountability for those actions.

Get started today

Ready to build your own trusted system of action? Start by exploring Spanner Graph to ground your agentic workflows in business reality. Next, use BigQuery Graph to build an auditable memory that powers closed-loop learning and regulator-grade explainability. You can begin capturing and analyzing these operational traces today using the BigQuery Agent Analytics Plugin and SDK. Finally, review the Ad Context Protocol to understand the open communication standards underpinning Yahoo’s agentic platform.

Modernizing Healthcare: How Alcidion achieved greater stability and performance with AlloyDB

Mon, 08 Jun 2026 16:00:00 +0000

In clinical informatics, every second counts. For Alcidion, a global leader in smart health solutions, the mission is simple but critical: use technology to reduce cognitive load for clinicians and present the right information at the right time to save lives.

Whether it’s managing patient flow in an emergency department or ensuring a patient is in the correct ward to avoid adverse outcomes, Alcidion’s flagship platform, Miya Precision, serves as a dynamic intelligent care platform for modern hospitals. To power this mission, the platform recently underwent a major architectural transformation, migrating from a legacy Microsoft SQL Server environment to Google Cloud’s AlloyDB for PostgreSQL.

The challenge: overcoming performance bottlenecks

Operating in an industry where data integrity and uptime are non-negotiable, Alcidion faced several technical and operational hurdles with its previous setup:

Operational overhead: Managing persistent backends for SQL Server required significant manual effort. The team had to manually balance database loads between elastic pools to maintain performance while trying to optimize costs. They also had to constantly manage the gap between allocated and used space to prevent shared pools from being consumed by excessive slack space.
Performance latency: Complex JSON data processing, critical for modern health informatics, was taking up to 30 minutes for certain jobs.
Stability concerns: The team sought a more stable Kubernetes environment and a persistent backend that could scale without constant administrative intervention.

The solution: a smooth migration to AlloyDB

Alcidion used the Database Migration Service (DMS) to move from SQL Server to AlloyDB, achieving a remarkably efficient cutover. The total learning and migration process took under one month, with the core database move completed in only one and a half weeks.

By creating custom synchronization tools and using Google Cloud’s managed services, the team reduced the final transition window to just 15 minutes. Alcidion achieved this by spinning up a new Google Cloud instance synchronized to the active one, with both accessible via unique fully qualified domain names. The new environment remained in read-only mode for customer validation.

During the final cutover, the old instance was set to read-only, synchronization was halted, and external integration links were toggled to the new environment. This streamlined process allowed users to log into the new instance and resume work within minutes, with the primary delay being DNS record updates.

Alcidion chose a fully managed AlloyDB service to eliminate control plane tasks and administrative overhead. This shift allows their engineering team to focus on clinical innovation and product development rather than "managing the container" or the underlying database infrastructure.

Being able to cut over to AlloyDB in about 15 minutes had our users back to work almost immediately. For a system clinicians rely on around the clock, that kind of smooth transition gave Alcidion real confidence.

The results: impact by the numbers

The shift to AlloyDB and Google’s Agentic Data Cloud has delivered immediate, quantifiable improvements for Alcidion and its healthcare customers:

Faster data processing: Data processing that previously relied on SQL Server stored procedures — a process that became increasingly time-consuming as data volumes grew — has been transformed. By migrating to AlloyDB and using BigQuery and Dataflow for processing, Alcidion has seen jobs that once took 30 minutes now complete in just 5 to 60 seconds.
Enhanced stability: The migration has delivered a step-change in reliability. In the previous environment, the team faced monthly disruptions, ranging from failed scheduled maintenance to connectivity issues that required manual intervention. In contrast, AlloyDB and Google Cloud’s compute services have proven exceptionally stable, allowing the team to move away from the "firefighting" mode associated with frequent infrastructure crashes.
Reduced cognitive load: By simplifying their backend and clinical dashboards, Alcidion’s SREs have significantly reduced their administrative burden. This shift has freed the team to focus on high-value innovation, such as refining predictive analytics and generative AI that empower clinicians to make informed clinical decisions faster.

Future vision: AI and beyond

Alcidion isn't stopping at database modernization. The move to AlloyDB is a foundational step for their next phase of growth:

AlloyDB columnar engine: The team is exploring the columnar engine for a second round of query optimization and real-time analytics.
Generative AI apps: Alcidion is actively working with Google to use AlloyDB’s Gemini Enterprise Agent Platform integration to perform concept analysis and pick out critical clinical insights from vast datasets.

By moving to AlloyDB, Alcidion has improved its stability and performance and built a strong foundation to keep delivering smarter, safer care to hospitals worldwide.

Ready to modernize your database? Learn more about how AlloyDB can transform your operational workloads.

How Trustpilot built a real-time architecture for data enrichment using Gemma

Mon, 01 Jun 2026 16:00:00 +0000

Processing millions of user reviews in real-time, under strict latency and cost constraints, is no easy task. Trustpilot has been doing exactly that with custom machine learning since long before large language models (LLMs) were cool. Now, as the company transitions its core stack to generative AI, here is a look at how we teamed up to build a high-volume streaming pipeline using fine-tuned Gemma models.

Powering deep review intelligence at scale

Trustpilot’s core business relies on delivering deep, actionable review intelligence. As a platform championing transparency and genuine feedback, it must safeguard data integrity and maximize value. This means extracting every drop of metadata from incoming reviews — making LLMs the perfect tool for the job.

These models excel at parsing messy, human-written text to run named entity recognition (NER), categorize business domains, score sentiment, and pinpoint customer intent. But while prompting an LLM for a few reviews is easy, processing millions in real-time without blowing up costs is a massive engineering hurdle.

Why fine-tune an open model?

When pursuing such a big task, why isn’t just plugging into a powerful, off-the-shelf, frontier model like Gemini the right approach? For a pipeline this critical to the core business, closed models are rarely the best option. Instead, by fine-tuning open-weight models like Gemma, Trustpilot takes full ownership of their AI strategy. Here’s how:

Total model independence: By owning its models, Trustpilot ensures it controls the retraining lifecycle, completely freeing it from a third-party vendor's update schedule or sudden API changes.
Predictable economics: Shifting from a variable per-token pricing model to fixed infrastructure costs makes running millions of predictions financially viable and optimizable.
Expanding MLOps capabilities: Building these models in-house enables Trustpilot to bake in the "secret sauce" of its review intelligence while building competencies on open-weight models.
Architectural continuity: Standardizing on an open-weight lineage preserves the company’s ability to leverage the future iterations of the base model. This enables performance gains with minimal engineering overhead.

Rather than deploying one massive model, Trustpilot built a suite of highly specialized models using the lightweight google/gemma-2-9b as a base.

To get heavy-weight performance from a small footprint, the company employed a consensus annotation over a stratified sample of the Trustpilot review corpus, using a selection of teacher models from the Gemini 2.0/2.5 Pro/Flash family. This process generated high quality training datasets for specialized tasks like topic classification, NER, and sentiment extraction.

The datasets were subsequently used to fine-tune a targeted lineup of custom models that considerably outperformed the legacy solution and delivered accuracy just a couple percentage points lower than the teacher models’ consensus.

System architecture

This architecture was built on top of Dataflow and Gemini Enterprise Agent Platform Endpoints, which play together very nicely because of the out-of-the-box VertexAIModelHandlerJSON.

We decoupled business logic and raw LLM inference by creating two separate endpoints:

The classifier: a FastAPI-based endpoint that handles the messy stuff, pre/post-processing, prompt templating, and chaining.
The LLM: A separate Agent Platform endpoint dedicated strictly to serving the Gemma model via vLLM.

This approach keeps the Dataflow job clean and ensures the LLM endpoint sticks to what it does best: generating text. Plus, it allows Trustpoint to scale them independently based on the traffic.

Performance tuning

To get the most out of the vLLM-based Agent Platform endpoints, Trustpilot focused on squeezing every bit of performance out of the entire pipeline, especially from the A2 VMs using A100 GPUs. It also leveraged the customized and optimized version of vLLM maintained by Gemini Enterprise Agent Platform.

A focus of our performance tuning involved optimizing the vLLM backend configuration to prevent processing bottlenecks. By carefully adjusting the engine parameters, selecting the appropriate data type, and enabling useful settings such as prefix caching, we ensured the models could smoothly handle high streaming volumes.

Together, we also created a reusable load testing framework to find the optimal serving capacity for a vLLM inference server and to sketch its performance profile. This enabled setting a baseline for needed infrastructure, and tuning the auto-scaling setup using the request count-based metric. In addition, a new metric using vLLM number of requests waiting could be even better for this.

Challenges

While building this setup, Trustpilot encountered a few notable hurdles:

Private networking: The architecture aimed to be fully isolated by using private endpoints and Private Service Connect, but this wasn’t possible because there was no native support for direct private communication between distinct endpoints.
Deployment observability and reliability: Endpoint deployments can be slow or opaque, which occasionally requires extra troubleshooting when entering an unhealthy state. Trustpilot is still working closely with the Gemini Enterprise Agent Platform product team to help shape future observability features and platforms.
GPU Scarcity: Securing A100 GPUs in the EU region is tough, so on-demand VMs are often a no-go. Instead, leveraging reservations is preferable but balancing them between development, production, training, inference, and experiments can be quite challenging.

The results

Together with Google Cloud, Trustpilot leveraged the full potential of Gemma on Gemini Enterprise Agent Platform to process millions of reviews a day in near real-time. In doing so, they achieved Gemini-like performance for a fraction of the cost. This ultimately allowed the Trustpilot Business Platform to turn millions of everyday customer reviews into instant, actionable insights. You can read more on the Trustpilot Medium blog post.

^{This blog post was written by Assulan Nurkas (Trustpilot), Subu Ramasubramanian (Trustpilot), Konrad Stanek (Trustpilot), Dario Banfi (Google) and Michael Cohen Hjertén (Google) based on the work done during the joint project at the end of 2025.}

Cool stuff Google Cloud customers built, May edition: Agentic algorithms for supply chains; virtual try-on APIs; robotic camera operators & more

Fri, 29 May 2026 16:00:00 +0000

AI and cloud technology are reshaping every corner of every industry around the world. Without our customers, who are building the future on our platform, there would be no Google

Cloud. In this regular round-up, we dive into some of the exciting projects redefining businesses, shaping industries, and creating new categories.

For our latest edition, we learn how Urban Outfitters sped up its order management; BASF uses AlphaEvolve algorithms to map global supply chains; the unification strategy for UKG’s workforce intelligence; WPP’s secrets to training humanoid robot camera operators; how Breuninger piloted Virtual Try-On APIs; creating automated video clips with Glance; and Movix improves the production of dental aligners.

Be sure to check back next month to see how more industry leaders and exciting startups are putting Google Cloud technologies to use. And if you haven’t already, please peruse our list of 1,302 real-world gen AI use cases from our customers.

Urban Outfitters saves big by migrating order management

Who: Urban Outfitters, Inc. (URBN), the popular clothing and home goods retailer, relies on IBM Sterling OMS as the nerve center of its global ecommerce operations. However, the foundation of this critical system — a massive 11TB Oracle database — was increasingly becoming a bottleneck.

What they did: URBN completed a major infrastructure upgrade, migrating its IBM Sterling OMS from an Oracle database to Google Cloud's AlloyDB for PostgreSQL. To enhance performance and provide high availability and scalability, the AlloyDB deployment architecture includes two read replicas, providing low-latency access to data for reporting and analytics. Google Cloud and IBM teams also assisted URBN in a rigorous, iterative switchover testing strategy.

Why it matters: The migration to AlloyDB has fundamentally reshaped URBN’s data strategy, delivering a more favorable total cost of ownership through an optimized storage and compute architecture, without sacrificing performance or reliability. Furthermore, the shift to a PostgreSQL-compatible database gave URBN the flexibility of an open-source ecosystem, providing freedom from vendor lock-in, as well as significant speed improvements that enhanced responsiveness.

Learn from us: "URBN’s successful migration serves as a blueprint for organizations looking to modernize their mission-critical infrastructure and future-proof their environment for AI expansion. This journey proves that even the most complex, mission-critical migrations can be achieved through deep cross-organizational partnership and a phased, risk-mitigated approach." – Rob Frieman, CIO, Urban Outfitters & Raj Pai, VP, Product Management, Databases, Google Cloud

BASF manages supply chain decisions with AlphaEvolve

Who: BASF Agricultural Solutions manages a complex network of 180 production sites with more than 5,000 distinct value chains. Currently, human planners make thousands of local decisions every day on what to produce, when to produce it, and how much safety stock to hold.

What they did: To understand how local decisions ripple across their entire global network, BASF turned to AlphaEvolve on Google Cloud to build a digital twin of their supply chain. In collaboration with Google Cloud and prognostica GmbH, BASF fed the model three years of historical data and then generated variations of the code, mutating the logic to see if it could simulate a supply chain that matched the real-world historical data.

Why it matters: By running thousands of experiments, AlphaEvolve developed a clear, human-readable algorithm that explains how the BASF network truly operates. The final algorithm successfully mirrored the actual historical performance of the supply chain, significantly reducing the error rates compared to the initial seed model. It automatically discovered factually correct, domain-specific supply chain rules, providing a clear foundation for optimizing asset utilization globally.

Learn from us: “We had several attempts to build a digital twin. … By using AlphaEvolve, we cannot only map the complex network based on system data, but at the same time understand and copy the human decisions that drive our daily operations.” – Dr. Goetz Krabbe, vice president for global supply chain at BASF

UKG unlocks real-time workforce intelligence at scale

Who: UKG is one of the leading providers of human capital management (HCM) and workforce management (WFM) solutions, but years of growth led to backend sprawl. They have 126 application teams, dozens of tech stacks, and more than 12,000 database instances.

What they did: To bring the full UKG suite onto one real-time foundation, the company built People Fabric, a new data and intelligence platform powered by AlloyDB for PostgreSQL and the just-announced Agentic Data Cloud. They created a custom change data capture (CDC) framework to extract changes from existing operational databases, and for larger analytical workloads, the same data flows into BigQuery, while Cloud SQL holds the metadata and tenancy context.

Why it matters: People Fabric gives UKG a complete and consistent view of people, work, pay, and culture data that’s updated continuously and ready for AI to use in real time. For engineering teams, People Fabric acts as a database-as-a-service that accelerates development and supports modernization without customer disruption. Additionally, migrating core person and employment data off their on-prem monolith has generated cost savings significant enough to fund half of People Fabric.

Learn from us: “As we continue expanding People Fabric, we’re laying the groundwork for deeper agentic automation, more responsive analytics, and a growing set of AI-driven capabilities — all on a trusted, scalable foundation built for what’s next.” – Radhi Chagarlamudi, Group Vice President, Product Engineering, UKG & Heather White, Cloud Data Architect, Google Cloud

WPP accelerates humanoid robot training 10x with G4 VMs

Who: WPP is one of the world’s largest marketing organizations, handling $70 billion of media for enterprise clients. They work on some of the most complex commercial film shoots and were eager to test the viability of robotic cameras to capture more footage, but this required complex training of physical models AI.

What they did: WPP used the new G4 VM instance powered by NVIDIA RTX PRO 6000 Blackwell on Google Cloud to tackle the unique challenges of training physical AI for robotics in videography settings. After capturing human motion with the OptiTrack mocap system, they undertook reinforcement learning using the AI Hypercomputer together with the NVIDIA Isaac Sim image. MuJoCo, an open source physics engine by Google DeepMind, was a critical piece of simulation software that validated accuracy continuously, in real-time.

Why it matters: WPP was able to utilize a P2P topology that moves data directly between GPUs without the bottleneck of central processing. They saw speed increases in excess of 10x, taking training times down to less than one hour. Through high-volume simulation, the humanoid robots learned how to respond to small changes and bridge the tough "sim-to-real" gap, helping ensure the robot's simulated adaptability translated to safety and stability in the real world.

Learn from us: "Our process for mastering complex, natural movement on a film set can be replicated across industries to overcome the massive computational complexity of training robots." – Perry Nightingale, SVP of Creative AI, WPP

Breuninger boosted sales with its "be your own model" AI

Who: Breuninger, a fashion and lifestyle company based in Germany, thought emerging generative media models could be a good fit to answer the question every online fashion shopper asks: "How will this look on me?"

What they did: Working with Google Cloud, they built a virtual try-on experience that lets shoppers see high-end fashion on their own bodies using a simple selfie. Using the Virtual Try-On (VTO) API, Breuninger’s data team worked directly with Google’s engineers to test and refine the technology in three stages, ultimately moving from pre-selected models to a user-first, selfie-based approach. The project was also part of Breuninger’s move to a Flutter-based platform, which helped the team move from its vision to a live launch in only three months.

Why it matters: During a six-week A/B test over Black Week and the holiday season, the team found that shoppers who used the virtual try-on converted purchases at a higher rate than those who didn't. Customer surveys reinforced the numbers: shoppers responded well to the high image quality and the personalized experience.

Learn from us: “Breuninger continues to refine the experience based on how customers actually use virtual try-on in everyday shopping — the same user-first approach that shaped the project from the start.” – Daniel Rascher, Senior Product Owner, Breuninger & Dr. Michael Menzel, Customer AI Specialist, Google Cloud

Glance turns hours of video into mobile-ready clips

Who: Glance, a mobile-first content platform, processes 1-2 hour videos from sources like podcasts, news reports, movies, and web series, and transforms them into 30 to 180-second vertical clips optimized for mobile lock screens.

What they did: The goal was to create a complete pipeline that takes a long-form landscape video (16:9) and outputs multiple ready-to-publish short-form portrait videos (9:16). The final technical solution uses Google Cloud Speech-to-Text v2, Gemini, and the Google Vision API, combined with custom video manipulation using Samurai (an open-source object tracking tool), OpenCV and MoviePy. The process involves audio extraction, speech-to-text transcription, and using Gemini 2.5 Flash to analyze transcript text and identify optimal start and end timestamps for short video clips.

Why it matters: With daily volume projected to grow from 3,500 to over 10,000 videos per day, manual editing wasn’t a realistic path forward. Glance’s video pipeline demonstrates what becomes possible when AI handles the repetitive, judgement-intensive work of video editing. The system transforms thousands of long-form videos into mobile-ready clips each day, preserving narrative context while optimizing for vertical viewing. Rather than choosing between scale and quality, automated pipelines can deliver both.Learn from us: “Glance’s video pipeline demonstrates what becomes possible when AI handles the repetitive, judgement-intensive work of video editing. … The approach offers a template for any organization sitting on long-form video archives. Rather than choosing between scale and quality, automated pipelines can deliver both.” – Himanshu Aggarwal,

Machine Learning Engineer, Glance & Sharmila Devi, AI Consulting Lead, Google Cloud

Movix fills a gap in dental skills with specialized agentic AI

Who: Movix is building one of the first agentic AI solutions for dental appliance manufacturers and dental labs, to help solve a serious shortage of skilled dental technicians in aligner manufacturing.

What they did: Movix developed custom models for deep learning, computer vision, and 3D mesh analysis over a five-month period, using Google Cloud infrastructure. Once defects are detected, they use the Gemini Enterprise Agent Platform to generate client-facing feedback that reads as if it came directly from a human technician. Their 3D models use Cloud Run with L4 GPUs for the massive compute power required, and they use Compute Engine VMs to run experiments and train models.

Why it matters: Movix’s agentic solutions automate data entry and quality control, which are traditionally manual, time-consuming, and error-prone tasks. The automation and higher level of accuracy the QC agent delivers can save $300 per remake for an aligner manufacturer, and speed up the appliance manufacturing process with quicker turnaround times.

Learn from us: “We plan to build hybrid solutions … designing an architecture that connects our cloud-based AI agents with older, on-premises software that many conservative labs still use — through lightweight local connectors and standardized APIs. This will allow us to access a large market segment that has not yet migrated to the cloud.” – Marina Domracheva, CEO, Movix & Bakit Dzhumagulov, CTO, Movix

Evolving Dataflow to process massive datasets for machine learning

Thu, 28 May 2026 16:00:00 +0000

Google created MapReduce more than 20 years ago to solve the scaling problems in data processing that the then young company was running into. The AI era that we are in now demands efficient, large-scale data processing for everything from training frontier models like Gemini by Google DeepMind to powering fully autonomous vehicles like Waymo.

Many aspects of machine learning, including data ingestion, transformation, and feature extraction, rely heavily on processing massive datasets. To meet this astronomical scale required by efforts across Google, we evolved our data platform, Flume, the successor to the original MapReduce, with innovations focused on scalability, efficiency, and a better developer experience. And many of those innovations are available as part of Dataflow, our fully managed batch and streaming platform built on the same core technology Google uses to power its most demanding internal workloads.. In this blog, we provide an overview of the many innovations in the Flume platform, and a glimpse into how Google Cloud customers are putting those features into action with Dataflow.

Addressing massive scalability

The scale of data processing at Google has exploded over the last 20 years and continues to drive innovation. To tackle the challenges of immense scale, we introduced several features within Google's data processing platform, which are also available in Dataflow::

Liquid sharding dynamically splits work units (shards) during execution for on-the-fly rebalancing. This helps pipelines with uneven data distribution and stragglers to maximize worker efficiency as data grows.
Global compute enables enormous scaling by dynamically scheduling workloads across Google's global infrastructure. The system automatically determines the optimal location based on factors like data locality and resource availability.
Automatic pipeline optimization fuses consecutive operations into a single stage. This reduces I/O and stage-transition overhead, allowing large-scale execution to scale more gracefully.
Rate-limiting external API calls manages load on external services. This is essential for modern ML pipelines that frequently call external APIs for tasks like model evaluation, preventing high data volumes from overloading systems.
Tandem pools facilitate serverless remote inference. This feature helps overcome scalability limitations often found in remote inference systems by efficiently hosting, sharing, managing, and autoscaling external model servers.

Boosting efficiency with accelerators

Doing more with less isn't just a constraint; it fuels our progress. By finding ways to run more efficiently, we create the space and capacity needed for rapid innovation. This is particularly evident for teams that use accelerators like TPUs for their workloads. To improve utilization and cost efficiency, our engineers devised several novel features for our platform, now part of Dataflow:

Heterogeneous worker pools allow developers to specify custom resource requirements for different pipeline stages. For example, TPU-intensive work runs on TPU-equipped workers, while other stages use standard CPU workers. This ensures optimal resource allocation.
TPU-aware autoscaling prevents excessive initial assignment of TPU workers and improves efficiency during subsequent autoscaling events.
Duty-cycle policy enforcement automatically scales down TPU workloads when the accelerator's duty cycle (the fraction of time it is active) is low, scaling back up only when utilization improves.
TPU fungibility: By working with other infrastructure teams, we developed optimizations to encourage scheduling jobs to the most suitable TPU version and cell location based on quota and resource availability.

Enhancing the developer experience

Considering the wide mix of backgrounds and tools across Google, rapid prototyping, iteration, and reliable production operations are extremely important. Google has invested in significant capabilities to enhance the overall user experience:

Language flexibility is provided through a versatile SDK with a simple API in C++ (internal to Google), Java, Python, and Go (with SQL support). This allows users to build batch, ML, and streaming pipelines.
Integration with ML frameworks like JAX is available, along with native support for LLM-specific optimizations. The underlying platform also provides building blocks for robust agentic inference pipelines and supports simple transitions between bulk and streaming paradigms.
Unified batch and streaming enables users to use the same code for both historical batch and live streaming data. This simplifies the architecture, which traditionally would have required separate pipelines for batch and streaming data processing.
Observability for production pipelines is available through the monitoring UI, which offers comprehensive control and essential diagnostic data. Detailed performance metrics, such as stage-level TPU utilization graphs, provide transparency for troubleshooting and optimization tasks.
Advanced developer workflows for quicker day 0 and day 2 operations include features like sampling and dry-run to help ensure code accuracy. Users can also test pipelines on small in-memory collections, and even pause and resume production pipelines.

Dataflow brings innovation from Google's internal platform to Google Cloud

Dataflow is built upon Google's internal platform, sharing many core components, including the execution engine and the Apache Beam SDK (which originated from Flume’s APIs). This close relationship means that the cutting-edge solutions we build to handle Google’s internal data processing challenges, like pipelines that process hundreds of billions of documents, directly benefit Dataflow users. In fact, unique Dataflow features like vertical scaling, right fitting, dynamic sharding, and straggler detection all resulted from solutions developed for Google’s internal workloads.

This is one of the reasons many Google Cloud customers rely on Dataflow for critical ML applications: Spotify uses Dataflow for large-scale generation of ML podcast previews. Etsy leverages Dataflow for data preparation and ETL for its ML workloads. And Moloco uses Dataflow to process terabytes of data a day to update its prediction model for real-time ad bidding.

The momentum continues: Last quarter we launched support for TPU in Dataflow in addition to supporting GPUs. Looking ahead, we are working on an advanced reliability feature called speculative execution and are enhancing the developer experience with features like failure isolation and replay and pause/resume, which are coming soon. To learn more or get started with Dataflow visit https://docs.cloud.google.com/dataflow/docs/get-started.

Announcing the newest cohort of the Google for Startups Accelerator: Middle East, North Africa & Turkey

Thu, 28 May 2026 07:00:00 +0000

Google’s mission is to organize the world’s information and make it universally accessible. In high-growth, technically ambitious markets like the Middle East, North Africa, and Türkiye (MENA-T), we fulfill this mission by supporting AI-First startups building the next generation of information-driven services on a global scale. In a region known for its resilience, we want to help founders flourish in any conditions.

The newest cohort of 15 companies in the Google for Startups Accelerator: MENA-T program starts on June 1. They follow on the success of our sixth group, which concluded in November 2025 and set a new benchmark for the region.

Over the course of the fall 2025 program, 14 AI-first startups from 8 different countries received more than 230 hours of specialized 1:1 mentorship from Google experts. This support allowed them to achieve measurable technical and business milestones, including refining their business strategies, accelerating AI/ML initiatives with Google Cloud, and enhancing overall product design.

We’re supplementing the 2026 program with additional resources, focus, and training to help these startups navigate the uncertain geopolitics that can affect the region and the world at any time.

Introducing the newest Google for Startups Accelerator: Middle East, North Africa, Turkey cohort

With a record breaking volume of applications, we are seeing more and more startups leveraging AI technology and addressing meaningful challenges with their business. Please join us in welcoming the 15 companies selected to participate in this cohort:

BioTwin creates virtual twins from health data to detect risks and recommend preventative actions.
Coral replaces manual sustainability processes with real-time enterprise overviews.
Each::labs builds the next generation of AI-native tools to streamline complex developer workflows.
Hakeem translates clinical studies into real-time, patient-specific guidance for clinicians.
inveon.ai deploys agentic AI to provide autonomous digital employees for e-commerce.
Jusoor Labs uses AI to analyze science experiment interactions and improve learning outcomes.
Openfarming automates distributor workflows to reduce waste and protect margins.
Plusfinity builds AI-native learning infrastructure for scalable, interactive education.
Promake empowers the manufacturing sector with AI-driven design and production optimization tools.
Qanooni transforms manual legal work into structured, searchable workflows.
Repzo uses AI to turn complex field data into natural language reports for field teams.
RFxAI streamlines procurement and sales through AI-driven response evaluation.
Tapper applies machine learning to detect anomalies and block invalid traffic.
TruBuild analyzes unstructured construction data for faster, objective tender evaluation.
Woliz uses voice AI to make digital ordering accessible for nanostore owners.

A curriculum designed for impact

Starting June 1st, founders will participate in a three-month program specifically tailored to help startups navigate their unique challenges. The curriculum provides intensive technical support, including comprehensive stack audits and one-on-one mentorship from global experts.

By balancing advanced technical training — focused on AI security and generative design — with strategic business modeling and go-to-market planning, we empower founders to scale their innovations securely. This holistic approach is designed to help startups maintain momentum and drive the region’s sustained digital growth and long-term resilience.

The program has already demonstrated significant impact for the fall cohort, with a number of startups accelerating their growth and development.

COGNNA, a provider of an agentic security operations center (SOC) suite, is among those seeing sustained growth. With improvements made during the accelerator, their platform now allows analysts to work 80% faster, and subsequently have closed a $9.2-million Series A funding round.

By using BigQuery to ingest petabytes of data and Google Kubernetes Engine to scale investigations, the startup has transformed its security operations and dramatically improved efficiency. "Google is shaping the future of COGNNA by enabling us to scale with global markets," said Ziyad Alshehri, co-founder and CTO of COGNNA.

Smart Bricks, a UAE-based startup for AI-powered real estate investing, recently closed a $5 million pre-seed round led by a16z Speedrun. Smart Bricks uses Google’s machine learning pipelines to automate 99% of manual real estate investment workflows across Dubai, London, and New York.

“The Google for Startups Accelerator played a key role in accelerating our technical development,” Mohamed Mohamed, founder and CEO of Smart Bricks, said. “Access to Google’s AI and cloud stack has been instrumental in building and scaling our agentic AI models, particularly given the scale and complexity of the data we’re working with. And infrastructure like Gemini Enterprise Agent Platform and BigQuery allowed us to significantly speed up our development cycles, improve model performance, and bring a much more robust, data-driven platform to market faster.”

Google’s commitment to MENA-T growth

We continue to support founders across the region, providing the specialized resources and cloud infrastructure needed to ensure that innovation continues to scale. Our goal is to ensure that the region’s digital economy continues its acceleration toward a more secure and innovative future.

We are excited to see how this new cohort will shape the future of the MENA-T ecosystem.

The Blueprint: How Movix fills a gap in dental skills with specialized agentic AI

Fri, 22 May 2026 16:00:00 +0000

Welcome to The Blueprint, a regular feature where we highlight how Google Cloud customers are tackling unique and common challenges across industries using the latest AI and cloud technologies. We hope to inspire others looking to innovate in their work.

The demand for dental appliances, like crowns and aligners, is booming, but it’s hard for manufacturers to keep up. At Movix, we’re building one of the first agentic AI solutions for dental appliance manufacturers and dental labs to help companies in the sector acquire digital technical expertise so they can scale clinical workflows cost-effectively and consistently.

The challenge:

Movix started in 2025 with a mission to solve a serious shortage of skilled dental technicians in aligner manufacturing through AI and agentic workflows. The need is significant: the global dental market is valued at nearly $400 billion and growing at double digits, yet many operations remain analog - creating enormous demand for co-pilot, agentic solutions.

Before founding Movix, we had previously started a vertically integrated dental aligner company that focused on very difficult dental situations, such as very crooked teeth. Yet even with highly skilled and trained technicians, there were often mistakes that would require remaking an aligner — a process that costs $300, roughly 25% of the retail price. Poor quality control took a real bite out of the company’s margins.

We saw an opportunity with Movix to address these mistakes by providing technicians with AI-powered quality control agents that automate aligner workflows and reduce errors. To achieve this, we needed to solve for a few key technical challenges:

Develop a custom AI model and end-to-end agentic workflow, since off-the-shelf solutions lacked domain expertise,
Ensure scalability would be built into the platform to prevent outages or production delays,
Achieve broad interoperability through a complex hybrid integration strategy since many dental practices are slow to adopt new technology and run on legacy systems.
Optimize security and compliance to comply with medical record regulatory requirements and keep patient data safe.

The solution:

In order to deliver AI agents that can provide expert-level accuracy, we needed to custom build a lot of the tooling ourselves. We started by developing our custom models for deep learning, computer vision, and 3D mesh analysis over a five-month period, using Google Cloud infrastructure. This intensive, methodical time helped ensure the right level of accuracy and quality control.

We use Google Cloud infrastructure across the full pipeline — from dataset storage and model training to evaluation — to build and refine our defect detection models for intraoral scans. Once defects are detected, we use Gemini Enterprise Agent Platform to generate client-facing feedback that reads as if it came directly from a human technician — acting as a digital team member in the quality control workflow.

Our 3D models use Cloud Run with L4 GPUs for the massive compute power we require; notably, performing the 3D segment scans and detecting defects across the entire fabrication process are highly compute-intensive processes. We use Compute Engine VMs to run experiments, along with various other GPUs to train our models, and perform the heavy lifting of model development in this environment.

Cloud Run and other tools like Cloud Storage support our scalability goals as we target large customers who handle high case volumes — some large labs might produce up to 200,000 appliances per year. Google Cloud's global network of data centers also simplifies regulatory compliance across regions and ensures fast delivery of large 3D datasets to clients worldwide.

The architecture:

The outcome:

Our agentic solutions automate data entry and quality control, which are traditionally manual, time-consuming, and error prone tasks. By automating the work of the best dental technicians, we’re ensuring a top quality product that will improve the fit of crowns, aligners, veneers, and implants for many, many patients. We estimate that our automation and the higher level of accuracy our QC agent delivers could save an aligner manufacturer $300 per remake, for example.

We also believe we’re helping to speed the appliance manufacturing process, leading to quicker turnaround times for dental appliances, which helps dental labs receive revenue faster and improve their cash flow. And we already know we’re meeting a critical need: After we launched the QC agent in October 2025, our first customer signed with us in December. That customer, Orthero, an aligner company serving more than 20 countries, has enjoyed significant results.

“Orthero benefits from this automation by making quality control faster, more consistent, and scalable,” Efer Turhan, a co-founder of Orthero, said. “With support from Movix’s QC AI Agent, we detect missing or inconsistent inputs early and flag unusual deviations before they cause delays.”

The details:

Even with the advantages of AI, our goals demand some serious work. Our architecture supports a solution that’s agentic and modular, integrates into existing on-premises dental systems, and ensures security and compliance.

Our agentic approach allows our system to run checks and balances, manage the complex, multi-step process of quality control for dental scans, and eliminate human errors that occur in data handling and quality review. Our goal is to develop five distinct AI agents by 2029 that cover the entire dental appliance workflow, from original patient dental scan to appliance manufacturing. While our first agents focus on data entry and dental scan quality control, our next agents will handle 3D file repair, clinical review, treatment planning, and manufacturing.

Our solution architecture also enables our system to integrate seamlessly with our customers’ existing lab management and manufacturing systems through API integrations. Because we are selling our solution into a conservative market, we decided to bear the burden of responsibility for successful adoption by doing as much of the integration work as possible.

Because we operate in the highly regulated healthcare industry, we built an environment that strictly follows compliance rules, anonymizing protected health information, or PHI, before it enters our machine learning pipeline to prevent health information from being exposed to the processing environment.

We plan to build hybrid solutions to capture a wider market as we move forward. We're designing an architecture that connects our cloud-based AI agents with older, on-premises software that many conservative labs still use — through lightweight local connectors and standardized APIs. This will allow us to access a large market segment that has not yet migrated to the cloud or begun to use new digital dental technologies.

Taken together, we are not just solving a skills gap, we are reimagining what is possible with co-pilot and agentic solutions across the entire dental industry.

How Glance turns hours of video into mobile-ready clips with AI

Thu, 21 May 2026 17:00:00 +0000

Every day, thousands of hours of new video content sits waiting to be discovered. Most of it lives in long-form, horizontal formats, while audiences are scrolling through vertical feeds on their phones.

Glance, a mobile-first content platform, knows this challenge well. The company processes 1-2 hour videos from sources like podcasts, news reports, movies, and web series, and transforms them into 30 to 180-second vertical clips optimized for mobile lock screens. With daily volume projected to grow from 3,500 to over 10,000 videos per day, manual editing wasn’t a realistic path forward.

The solution also needed to go beyond simple cropping. It required the intelligence to identify and center the primary speaker, or dynamically split the screen to stack speakers vertically during conversations, preserving the context that makes content worth watching.

Here’s how Glance’s video generation solution works.

Building for the lock screen era

The goal was to create a complete pipeline that takes a long-form landscape video (16:9) and outputs multiple ready-to-publish short-form portrait videos (9:16). The solution needed to handle:

Key Moment Identification: Finding the most engaging 60-second segments within hours of long-form footage
Active Speaker Detection: Identifying who’s talking in each frame and positioning them at the top of a split screen. This includes distinguishing between a static image and a live person to ensure the crop focuses on the actual speaker.
Split Screen Detection: Recognizing interview layouts (common in news broadcasts) and stacking the frames vertically to preserve conversation context
Intelligent Reframing: Converting a multi-speaker, wide-screen shot into a focused, vertical frame without losing context
Dynamic Caption Highlighting: Generating word-level timestamps for "Karaoke-style" captions that increase engagement on silent-by-default mobile screens
Automated Branding: Applying masks, logos, and overlays programmatically to maintain brand consistency across all videos

The final technical solution uses Google Cloud Speech-to-Text v2, Gemini, and the Google Vision API, combined with custom video manipulation using Samurai (an open-source object tracking tool), OpenCV and MoviePy.

Architecture overview

The pipeline is divided into three distinct modules.

Fig. 2: High-level architecture

Module 1: Video clipping

This module converts long videos to transcripts, identifies key segments, and clips the video. Accuracy matters here: precise word-level timestamps ensure clips start and end exactly where they should.

Fig. 3: Video Clipping Workflow

The process involves audio extraction, speech-to-text transcription, and timestamp identification using generative AI. The module performs the following key functions:

Audio extraction: Extracting the audio from the original video file.
Speech-to-text transcription: Converting audio into text with precise timestamps for each word
Segment identification: Using Gemini 2.5 Flash (aka Nano Banana) to analyze transcripts text and identify optimal start and end timestamps for short video clips
Video clipping: Clipping the video into short segments based on the identified timestamps
Transcript validation: Using Gemini to verify phrases and words are accurately captured (this step does not validate word timing)

The output is a set of short video clips, each paired with its time-aligned transcript, ready for the next stage: the Intelligent Reframing Engine.

Module 2: Intelligent Reframing Engine

The core technical work here is converting a horizontal 16:9 frame into a compelling 9:16 vertical frame. A simple center crop often cuts out key speakers or action, so our solution uses a multi-stage scene analysis pipeline.

Fig. 4: Intelligent reframing engine

Active speaker detection

To know what to crop, we first need to know who’s talking. This happens on a frame-by-frame basis using the face detection capabilities of the Google Cloud Vision API.

Fig. 5: Active speaker detection

The liveness check: Differentiating a live speaker from a static image (like a photo on the wall or a graphic) is essential. This was achieved by tracking facial landmarks:

Mouth movement: Calculating the normalized distance between upper and lower lip landmarks
Head movement: Tracking changes in head pose angles (pan, roll, tilt)
A face must show consistent animation in these cues to be classified as a "live" participant

Quantifying engagement: Once confirmed as live, we calculate an activity score based on:

Mouth openness
Emotional fluctuation (changes in joy, surprise, etc., provided by Vision API)

Primary speaker identification: The final decision uses a liveness ratio:animated frames divided by total frames where the face appears. The person with the ratio closest to 1.0 (meaning they were consistently animated on screen) is designated as the primary speaker.

One edge case addressed during the development was a static background image appearing behind a live news anchor (as shown in Fig. 6). The liveness check handles this correctly because the static image shows no facial animation.

Fig. 6: Scenario with one active speaker and one static background image

Split-screen detection

This step addresses interview scenarios where two subjects appear on opposite sides of the landscape frame. The system detects split-screen layouts and stacks the two halves vertically to maintain conversation context.

Fig. 7: Video reformatting

With active speaker detection complete, the system uses the primary speaker's location to identify split-screen segments. The goal is to find the precise dividing line between panels, enabling the video to be reformatted into a vertical, top-and-bottom layout. Two complementary approaches accomplish this:

Approach 1: Continuous face tracking with Samurai

This method uses Samurai, an open-source object tracking tool, to follow the primary speaker continuously. The trajectory is analyzed for split-screen layouts based on:

Consistent off-center positioning: The speaker remains on one side of the screen (e.g., left or right half), indicating a split panel rather than free movement across the frame.
Vertical dividing line detection: Image analysis identifies a persistent vertical line separating the two panels.
Background discontinuity analysis: Differences in color, texture, and scenery between the speaker’s background and the opposite side confirm two separate video feeds.

Fig. 8: Background discontinuity analysis

Approach 2: Frame-by-frame detection with Google Cloud Vision API

This approach uses Cloud Vision API's face detection to identify split-screen layouts based on the primary speaker's face location:

Off-center face: Consistent face detection in one region (such as the left 40% of the frame) flags a potential split screen.
Proximate dividing line: Vertical lines between the face and the screen center confirm a panel boundary.
Contrasting backgrounds: Inconsistent backgrounds between the speaker's side and the far side confirm the split-screen layout.

The output: Vertical stacking

Once the system recognizes a split-screen, it performs a digital cut-and-paste. This preserves both speakers and their reactions in a mobile-native format.

Automated reformatting

With the scene analysis complete, the OpenCV2-based solution intelligently applies the appropriate reframing rule to each segment:

Single speaker crop: For scenes with one primary speaker, the system anchors the 9:16 frame to the speaker’s face, keeping them centered.
Split screen: When a split is detected, the system slices the frame along the dividing line and stacks the panels vertically (left panel on top, right panel on bottom).
Multi-speaker crop: For scenes with multiple people (not a formal split), the system focuses the crop on the most prominent speaker or the face closest to the center.
Fallback: If no faces are detected (e.g., graphics or wide shots), the system applies a center crop or horizontal padding (letterboxing).

Two final techniques ensure a polished look:

Short scene merging: Segments shorter than a defined threshold merge with the preceding or following scene, eliminating flicker.
Camera smoothing: When focus shifts between speakers, a virtual camera effect creates a slow pan from one position to the next, rather than an abrupt cut.

Module 3: Finishing and branding

The final stage ensures the clips are ready for immediate publication, focusing on viewer engagement and brand reinforcement.

Dynamic caption highlighting

Using the word-level timestamps from the speech-to-text module, the system overlays highlighted captions with MoviePy. This involves:

Fig. 9: Dynamic caption highlighting

Sentence reconstruction: Grouping individual words into readable lines that adhere to character limits
Highlighting: The currently spoken word is highlighted in a distinct color (mustard yellow) against a black background, a proven method for increasing engagement when videos play without sound.

Masking and logo placement

Two overlay techniques maintain consistent branding across all videos:

Mask placement: A PNG mask with an alpha channel resizes the video to fit precisely into the transparent area. The mask's opaque regions (such as colored bars) serve as a dedicated background for captions and persistent graphics.
Logo overlay: The brand logo is placed onto the video based on configurable parameters for position (top-right, bottom-left, and so on), size, and margin.

Fig. 10: Mask and logo placement

Conclusion

Glance’s video pipeline demonstrates what becomes possible when AI handles the repetitive, judgement-intensive work of video editing. By combining speech-to-text transcription, computer vision, and generative AI, the system transforms thousands of long-form videos into mobile-ready clips each day, preserving narrative context while optimizing for vertical viewing.

The approach offers a template for any organization sitting on long-form video archives. Rather than choosing between scale and quality, automated pipelines can deliver both.

If you’re exploring similar video processing, content transformation, or media AI projects, the Google Cloud consulting team is eager to connect and explore the possibilities. For more on the AI products used in solutions Glance’s this, visit our AI & ML Products page.

_{This solution was a collaborative effort between Glance ( Pradeep Tiwari , Himanshu Aggarwal) and Google Cloud Consulting (Sharmila Devi, Jinyeong Yim, Rohit Sroch, Neeraj Shivhare and Kinjal Singh).}

Urban Outfitters achieves major cost savings by moving Sterling OMS to AlloyDB for PostgreSQL

Wed, 20 May 2026 16:00:00 +0000

Editor’s note: Urban Outfitters, Inc. (URBN) recently completed a major infrastructure upgrade, migrating its IBM Sterling Order Management System (Sterling OMS) from an Oracle database to Google Cloud's AlloyDB for PostgreSQL. This strategic move, a testament to the growing partnership between Google Cloud and IBM, delivers significant benefits for URBN, paving the way for increased efficiency, reduced costs, and a future-proofed technology landscape. This success story showcases how businesses can leverage AlloyDB for PostgreSQL to modernize their databases and unlock new levels of performance and scalability.

In the fast-paced world of retail, order management is the backbone of a seamless customer experience. Urban Outfitters, Inc. (URBN) relies on IBM Sterling Order Management System (Sterling OMS) as the nerve center of its global ecommerce operations, orchestrating everything from order capture and real-time inventory tracking to fulfillment optimization and post-purchase logistics. This system helps ensure that URBN can efficiently process millions of transactions across its global network of stores, warehouses, and digital channels, delivering on customer expectations for fast, accurate, and flexible order fulfillment.

However, the foundation of this critical system — a massive 11TB Oracle database — was increasingly becoming a bottleneck. High licensing and maintenance costs, growing operational complexity, and the constraints of proprietary technology posed significant challenges to scalability and long-term innovation. To maintain Sterling OMS's high availability, performance, and transactional integrity, URBN needed a modern database solution that could:

Reduce total cost of ownership (TCO): Lower licensing, operational overhead, and infrastructure expenses while maintaining reliability.
Ensure business continuity: Support high availability, rapid failover, and disaster recovery to prevent disruptions in order processing and customer transactions.
Embrace open standards: Transition from proprietary technology and embrace open, flexible, and future-proof solutions.
Maintain feature parity: Ensure a seamless migration without disrupting Sterling OMS functionality, keeping all mission-critical capabilities intact.

For a retail enterprise like URBN, even minor disruptions to order management can have significant financial and operational consequences. A failed transaction, an inventory miscalculation, or a delay in fulfillment can directly impact customer satisfaction, brand reputation, and revenue. Because Sterling OMS is so mission-critical, URBN required a migration approach that was as technically robust as it was precise — demanding a transition with near-zero downtime, data loss, or performance degradation.

The solution: AlloyDB for PostgreSQL

The success of this complex transition hinged on a deep, ongoing collaboration between URBN, IBM, and Google Cloud. This partnership brought together industry-leading expertise and cutting-edge technology, with teams working in lockstep to ensure high-touch engagement throughout every phase. By embedding dedicated IBM and Google Cloud engineers directly with URBN’s technical staff, the teams were able to meticulously plan and optimize the migration of the massive database.

The project’s success was defined by several critical pillars:

First-tier database recognition and feature development: IBM and Google Cloud engineering teams collaborated to ensure that Sterling OMS fully recognized and supported AlloyDB for PostgreSQL as a first-tier database.
Enterprise-grade reliability with two read replicas: To enhance performance and provide high availability and scalability, the AlloyDB deployment architecture includes two read replicas, providing low-latency access to data for reporting and analytics and improving operational resiliency of the entire Sterling OMS application.
Extensive performance tuning: A dedicated performance engineering team from Google Cloud worked alongside URBN and IBM experts to fine-tune database queries and optimize configurations. This level of continuous, high-class support ensured AlloyDB not only met but exceeded the performance benchmarks of the previous Oracle database. This was essential to handle the high transaction volume of the Sterling OMS on AlloyDB for a very large retail customer, the size of URBN.
Rigorous switchover testing and risk mitigation: Google Cloud and IBM teams assisted URBN in a rigorous, iterative switchover testing strategy, which involved running the Sterling OMS system on AlloyDB for a full day before switching back to the Oracle database. This proactive testing allowed URBN teams to identify and resolve potential issues in a controlled environment, significantly reducing risks and increasing confidence in the migrated system.

aside_block: <ListValue: [StructValue([('title', 'Build smarter with Google Cloud databases.'), ('body', <wagtail.rich_text.RichText object at 0x7fdc31dc55e0>), ('btn_text', ''), ('href', ''), ('image', None)])]>

A transformative shift

The migration to AlloyDB has fundamentally reshaped URBN’s data strategy, delivering a more favorable TCO through an optimized storage and compute architecture, without sacrificing performance or reliability. Furthermore, the shift to AlloyDB, a PostgreSQL-compatible database, gave URBN the flexibility of an open-source ecosystem. This move not only provides freedom from vendor lock-in, but also connects URBN to a vibrant community and a vast array of modern tools, ensuring long-term technical agility.

Beyond cost and flexibility, the transition unlocked superior performance and scalability to support URBN’s mission-critical operations. The combination of an optimized database kernel and precise query tuning resulted in significant speed improvements, directly enhancing the responsiveness of the Sterling Commerce system.

A blueprint for success: Planning and testing

URBN’s successful migration of IBM Sterling OMS to AlloyDB serves as a blueprint for organizations looking to modernize their mission-critical infrastructure and future-proof their environment for AI expansion. This journey proves that even the most complex, mission-critical migrations can be achieved through deep cross-organizational partnership and a phased, risk-mitigated approach.

For any enterprise navigating the challenges of modernization, URBN’s experience offers a clear roadmap for success. The use of iterative switchover tests — running the system on AlloyDB and switching back — was the "secret sauce" that built the necessary confidence for the go-live. By prioritizing this level of rigorous testing, businesses can move toward a future of greater agility, efficiency, and innovation.

Learn more:

Discover how AlloyDB combines the best of PostgreSQL with the power of Google Cloud in our latest e-book.
Try AlloyDB at no cost for 30 days with AlloyDB free trial clusters!
Learn more about AlloyDB for PostgreSQL.

How Imgix processes 8 billion images daily with G4 VMs powered by NVIDIA Blackwell

Tue, 12 May 2026 16:00:00 +0000

The modern web is extremely visual. People are busy and easily-distracted, and smart companies know they have just seconds to attract would-be customers with compelling images, videos, animations, and other eye-catching elements. That’s why iconic brands like Bugatti, Yeti, Porsche, Spotify, and Sonos rely on Imgix to be the engine driving their online visual media.

Every day, Imgix serves more than 8 billion images and videos for brands like these and many others. With a platform designed to unify media optimization, AI transformation, and global delivery, Imgix ensures that its partners’ digital experiences are fast, personalized, and built for performance. Now more than ever, leading organizations are demanding real-time, high-fidelity media, and they need it to be fast.

To meet that demand, Imgix has evolved its infrastructure from private data centers to a full-stack, GPU-based environment on Google Cloud’s AI Hypercomputer. By transitioning to G4 VMs powered by NVIDIA RTX PRO 6000 Blackwell GPUs, Imgix ramped up its real-time processing capabilities, cutting median latency by 50% and increasing throughput per node by 6x. And it did all of that without changing its core application code.

The challenge: Instant visuals at scale

To capture people’s attention businesses need rich, fast-loading content that can reach millions of users simultaneously across a diverse array of devices.

A big part of that is real-time transformations — resizing, format negotiation, and applying artistic effects — and the computational power required for real-time transformations can be immense.

With inefficient technology, load times can be slow and brands risk giving their users poor experiences. Imgix’s solution to this challenge is a "just-in-time" philosophy. Achieving this requires high-performance instances. And with G4 VMs, they were able to process images instantly upon request rather than pre-rendering and storing millions of image variations.

Adopting the system that runs Google

When companies build on Google Cloud, they get more than just servers: they plug into the same intelligence engine powering Google's many billion-user products. Imgix is leveraging this structural advantage by using G4 VMs.

G4 VMs incorporate eight NVIDIA RTX PRO 6000 Blackwell GPUs, two AMD Turin CPUs, and Google Titanium offloads, which act as a dedicated administrative assistant for businesses’ servers. They handle the ”office chores” of security and data traffic in the background while the main processor does a company’s heavy lifting.

The G4 VM’s custom P2P interconnect yields up to 168% more throughput than standard configurations. With this architecture, Imgix can move all its image processing operations to NVIDIA GPUs and run multiple requests in parallel.

Inside the Imgix architecture

Imgix offers more than 150 different visual filters and its architecture is built to handle transformation requests dynamically based on which filters users choose. The pipeline has four primary stages:

Ingestion: The system retrieves assets directly from customers and routes them to a 2.5 petabyte storage cache on Google Cloud Storage (GCS). This high-speed layer replaces unreliable random web requests with a redundant, geographically distributed infrastructure.
Decoding: High-performance C libraries, supplemented by nvJPEG, decode assets into raw RGBA data. This leverages the G4 VM’s massive parallelism to handle multiple decoding stages, including Huffman decoding, Inverse DCT, and color space conversion.
Transformation: A custom Vulkan compute shader stack handles the core processing. Instead of fixed graphics pipelines, these shaders treat transformations (like resizing or masking) as parallel math problems rather than standard graphics tasks, enabling thousands of simultaneous pixel operations on the G4 VM clusters.
Encoding and Delivery: Once transformed, images are re-encoded using hardware-accelerated tools like NVENC and delivered via a global CDN. Because the G4 VM includes independent hardware engines for NVENC (encoding) and NVDEC (decoding), concurrent image manipulations on the CUDA cores aren’t slowed down.

Advanced video and image intelligence

Imgix is also using NVIDIA’s CUDA libraries for high-performance video analytics. By integrating NVIDIA DeepStream, it executes real-time object tracking within video streams for automated content analysis.

For static imagery, meanwhile, Imgix uses the nvJPEG library to offload computationally intensive JPEG decoding directly to the GPU. This prevents CPU bottlenecks during the ingestion of high-resolution assets while allowing the custom Vulkan compute shaders to begin immediate pixel-level transformations on the raw RGBA data residing in GPU memory.

The results: 50% faster and up to 6x more throughput

Thanks to its transition to G4 VMs, Imgix achieved the significant performance gains mentioned above without having to rewrite its core logic:

A 50% reduction in processing latency: It cut median latency from 100 milliseconds to 50 milliseconds.
A 5x to 6x increase in throughput: Its G4 VMs now handle up to six times the workload of its previous generation nodes.
Seamless migration: Imgix supported the G4 VMs by updating its Terraform scripts without needing to implement any application code changes.

"Building on Google Cloud's AI Hypercomputer isn't just about optimizing our current workloads; it's about future-proofing our platform. It gives us the foundational power to seamlessly weave advanced generative AI capabilities into real-time workflows, allowing our customers to push the boundaries of visual storytelling at global scale." - Alfonso Acosta, Head of Engineering, Imgix

Orchestrating at scale

To support the billions of image and video requests its customers process every day, Imgix built a sophisticated hybrid orchestration model:

Management: Google Cloud Run manages session and account layers.
Core Processing: Google Compute Engine-managed instance groups host the G4 VMs, which allows custom software to use the entire machine with no container "slicing."
Dynamic Scaling: Autoscaling relies on custom application metrics, such as machine queue length, rather than standard CPU use. This ensures that the stack’s most expensive elements are tuned for maximum efficiency.
Self-Healing: A custom mechanism monitors logs for driver faults, automatically "reaping" and restarting GPU instances without manual intervention.
Optimization: To maintain peak performance, Imgix uses NVIDIA Nsight Systems to identify and resolve code bottlenecks.

The future: From experimentation to execution

Even with the significant performance improvements it’s already achieved, Imgix is continuing to expand its AI infrastructure so its customers can access additional advanced capabilities like generative fill, background replacement, object removal, and image upscaling.

Features like these rely on high-performance machine learning systems that must process increasingly complex computations with no loss of speed or quality. By leveraging Google’s AI Hypercomputer, Imgix is now deploying and serving these models efficiently and offering its customers real-time, production-ready AI editing. And as demand grows for more dynamic and personalized visual experiences, this foundation is ensuring that Imgix can continue to deliver powerful capabilities reliably and at scale.

Get started

G4 VMs work natively with Google Compute Engine, Google Kubernetes Engine, Google Cloud Storage, and Vertex AI.

Dive deeper: Explore the Imgix architecture on GitHub.
Start building: Read the G4 VM documentation.

How BASF manages thousands of supply chain decisions with AlphaEvolve’s agentic algorithms

Thu, 07 May 2026 16:00:00 +0000

The agricultural and crop protection supply chain is one of the most intricate networks in the world. It takes up to two years to turn active ingredients into the final products farmers need, and a single change in weather or regulations can disrupt everything. Planners at BASF Agricultural Solutions navigate this reality daily across 180 production sites. To understand how local decisions ripple across their entire global network, BASF turned to AlphaEvolve on Google Cloud to build a digital twin of their supply chain.

Planning across a two-year lead time

BASF Agricultural Solutions manages a network with over 5,000 distinct value chains. Creating a single end product requires a bill of materials that can be over 30 levels deep, moving across different production sites and regions.

Currently, human planners make thousands of local decisions every day. They decide what to produce, when to produce it, and how much safety stock to hold. Because the network is so large, a planner can’t easily see how a localized decision affects the rest of the global supply chain.

This scale can lead to additional working capital and inventory and or cause production imbalances. Traditional mathematical models struggle to capture the dynamic reality of the network that planners navigate based on years of experience.

Building a foundation for decision support

AlphaEvolve is an evolutionary coding agent that generates and refines algorithms autonomously. In collaboration with Google Cloud and prognostica GmbH, BASF’s objective was not to replace human decision-making, but to establish a new model for decision support that helps planners handle the real-world complexity of the production network.

The team gave AlphaEvolve a foundational "seed" program. This initial code established a standard planning logic that translated demand forecasts into production schedules, serving as a functional baseline before introducing dynamic, network-wide coordination. From there, they fed the model three years of historical data, including inventory levels, market demand, and actual production outputs. AlphaEvolve then generated variations of the code, mutating the logic to see if it could simulate a supply chain that matched the real-world historical data.

Measuring what good looks like in initial tests

For AlphaEvolve to improve, it needed a specific goal. The evaluation function scored every new piece of generated code on one primary metric: how closely the simulated inventory levels and production decisions matched the actual historical reality recorded by BASF.

The latest AlphaEvolve runs delivered more than 80% relative improvement in accuracy compared to the initial seed model. With further adjustments, the team expects to push performance even higher — bringing the model to a level of accuracy not achieved with other approaches and making it actionable for operational use.

The results

The evolved planning logic delivered immediate, measurable improvements over the initial seed model. The final algorithm successfully mirrored the actual historical performance of the supply chain, significantly reducing the error rate compared to the initial seed.

“We had several attempts to build a digital twin for our complex supply network using deterministic models, and all of them failed,” said Dr. Goetz Krabbe, vice president for global supply chain at BASF. “By using AlphaEvolve, we cannot only map the complex network based on system data, but at the same time understand and copy the human decisions that drive our daily operations. This gives us a highly accurate and easy to maintain data driven digital twin of the entire network. Using it we can optimize our inventory levels and respond to market volatility with confidence while avoiding stockouts."

What the evolved algorithm actually does

By running thousands of experiments, AlphaEvolve developed a clear, human-readable algorithm that explains how the BASF network truly operates. It automatically discovered factually correct, domain-specific supply chain rules that explain the observed production outputs and inventory levels for the tested product value chain:

Production consolidation: The algorithm learned to group production amounts together, accurately mapping how planners optimize plant time.
Dynamic safety stocks: It introduced safety stock parameters to handle volatile and seasonal demand patterns, helping to strictly manage capital costs while preventing out-of-stock situations.
Network-wide coordination: The model successfully mapped the dependencies between different production tiers, providing a clear foundation for optimizing asset utilization globally.

What's next

The initial simulations showed that evolutionary AI can accurately model large-scale, dynamic supply chains. BASF’s objective is to create a digital twin of their entire global production network as a new foundation for simulation, decision support, scenario forecasting and optimization. This will allow the team to continuously simulate operations, identify hidden bottlenecks before they affect throughput, and optimize asset utilization across all global facilities.

_{This project was a collaboration between the BASF SE team including: Benjamin Priese, Michael Arlt, Debora Morgenstern and Tobias Hausen as well as Manuel Doerr and Thomas Christ from Prognostica GmbH Würzburg, and the AI for Science team at Google Cloud including (but not limited to): Kartik Sanu, Laurynas Tamulevičius, Nicolas Stroppa, Chris Page, Srikanth Soma, John Semerdjian, Skandar Hannachi, Vishal Agarwal and Anant Nawalgaria as well as Christoph Tittelbach from the Google account team and partners at Google DeepMind}

Fitting the future: How Breuninger boosted sales with its "be your own model" AI

Wed, 06 May 2026 16:00:00 +0000

“How will this look on me?” It’s the question every online fashion shopper asks, and one that most retailers still can’t answer well.

Breuninger, a fashion and lifestyle company based in Germany, thought emerging generative media models could be a good fit for this fashion conundrum. Working with Google Cloud, they built a virtual try-on experience that lets shoppers see high-end fashion on their own bodies using a simple selfie.

From trusted tester to live product

The project began when the Google Cloud team in Germany invited Breuninger to join the Trusted Tester Program for the Virtual Try-On (VTO) API. Breuninger’s data team in Germany worked directly with Google’s engineers in California, testing and refining the technology in three stages:

Catalog enrichment: The team first explored the VTO API to dress professional models in different outfits. This helped Breuninger to cover a greater variety in user tests without having to plan new photoshoots.
Body type selection: They then added a feature that let users choose from different body types to see how clothes would drape on a silhouette similar to their own.
The 'Be your own model' breakthrough: User feedback showed that customers did not just want to see a model; they wanted to see themselves.

The product owner at Breuninger noted that this close collaboration allowed the team to share user feedback with developers in real time. This speed helped them move from using pre-selected models to a user-first, selfie-based approach.

Three levels of virtual try-on

The project revealed three levels at which retails can adopt VTO, depending on how much personalization they want to offer:

Approach	Interaction	Use case
Level 1: Catalog enrichment	Offline batch processing	Dress standard models in new collections at scale to update product pages without manual shoots.
Level 2: Body type selection	Online on-request	Offer predefined models for users to choose from, similar to the virtual try-on feature on Google Shopping.
Level 3: 'Be your own model'	Online personalized	The most personal experience where users upload a selfie to see themselves in specific items or full outfits.

Building for scale with Flutter

Scaling a personalized experience required more than just an AI model. Selfies come in wildly different lighting and quality, so the team built preprocessing tools to make sure the final images met Breuninger’s brand standards. This project also accelerated Breuninger’s move to a Flutter-based platform. The VTO feature was the first module built by a self-sufficient product team using this new structure, which helped the team move from a vision to a live launch in only three months.

Real results during the holiday season

During a six-week A/B test over Black Week and the holiday season, the team found that shoppers who used the virtual try-on converted at a higher rate and generated a stronger contribution margin than those who didn't. Customer surveys reinforced the numbers: shoppers responded well to the high image quality and the personalized experience. Perhaps most telling, the team found that VTO became a tool for building style confidence — helping customers feel sure about a purchase before they made it.

What’s next

The pilot’s success has set up a broader rollout and international expansion, with physical fit and sizing support on the roadmap. Breuninger continues to refine the experience based on how customers actually use it in everyday shopping — the same user-first approach that shaped the project from the start.

To explore how generative AI can help your business create similar experiences, visit Google Cloud's Virtual Try-On solution. You can also try the feature yourself in the Breuninger app.

_{This work wouldn’t have been possible without the contributions from peers at both Breuninger, and Google Cloud. Thanks to Markus Peetz, Jorina Hilser, Martin Csengeri, Jay Deutinger, Sofia Widmayer, David Schowalter, Tobias Götze, Eric Karge, Abdul Mateen, Besnik Brahimi, Oliver Fesseler, and Lisa Beutner from Breuninger, and Khanh LeViet, Jorj Ismailyan, and Matt Chaban from Google Cloud.}

The Blueprint: Translating stream-of-conscious speech into responsive, actionable task lists

Wed, 06 May 2026 16:00:00 +0000

Welcome to The Blueprint, a new feature where we highlight how Google Cloud customers are tackling unique and common challenges across industries using the latest AI and cloud technologies. We hope to inspire others looking to innovate in their work.

Founded in 2007, Doist is a pioneer in async and remote-first work on a mission to simplify life’s complexities through apps like Todoist for task management and Twist for team communication.

The challenge:

We launched Ramble to take our popular Todoist application to the next level by capturing non-stop, stream-of-consciousness talking. Our inspiration was that scene from The Devil Wears Prada where Miranda Priestly rapid-fires a dozen tasks at her assistant. We asked: What if anyone could capture tasks that way? No typing, no careful formatting. Just talk and let Todoist do the organizing. That use case became our north star.

At the outset, we identified four big technical hurdles:

We needed fast and accurate real-time communication with tool-calling capabilities.
Multilingual support at scale but with great support for slang, accents, and more.
As traditional assertion-based testing would not work for our platform, we would have to find a way to achieve non-deterministic output testing and semantic validation.
Reliable, flawless handling of audio across browsers.

The solution:

We built Ramble using Gemini Enterprise Agent Platform and its previous iteration, Vertex AI; specifically, we’re using Agent Platform to access the Gemini Flash models. We chose these over other options primarily due to the quality of Google's state-of-the-art models and its clear terms and assurances about preserving privacy.

Gemini’s Live API (accessed via Agent Platform) powers Ramble’s core real-time interactions and key capabilities, including native audio streaming, proactive tool calling, session resumption, and multilingual understanding.

Ramble sends the raw PCM audio directly to the model without pre-transcription. Gemini handles language detection, speech recognition, and semantic understanding in a single pass, reducing latency. It then invokes our purposefully designed tools (addTask, editTask, deleteTask, etc) autonomously as the user speaks, without waiting for explicit commands.

The APIs in Agent Platform provide resumption tokens that let users pause and continue sessions, which is essential for mobile users who might switch apps or lose connectivity.

The end result is a clear, concise list of the tasks, regardless of how many, how inconsistently, or how confusingly they may have been rambled by the user.

The architecture:

The outcome:

Ramble has come to rely on the quality of Google’s AI models, particularly the reasoning and near-instant audio-processing capabilities of Gemini Flash. Other platforms and models offer similar capabilities, and we did bake in support for them, but none hit our internal quality bar as consistently as Gemini. When it came to a user's unstructured “rambling” and the need to fill in gaps, Gemini turned out to be the most intelligent of all the models we explored. The result was the clearest and most consistent breakdown of tasks, which was the exact magical user experience we wanted to create.

After an early rate-limit incident caused by unexpectedly high usage during alpha testing, we developed a deeper, more proactive partnership with Google, ensuring long-term sustainability and the support necessary for our high API usage. Since then, it's been easy for us to connect directly with Google Cloud staff, including engineers, when issues arise.

Here at Doist, Ramble took off both in a qualitative and quantitative sense. It’s become a hallmark experience that incentivizes us to explore tasteful applications of AI that can enhance our existing product experience, both in the B2C space as well as B2B. Beyond task creation, we’re considering several opportunities across the productivity journey, from capture to planning and even automation.

The details:

We structured our back-end to enable future voice-powered features. The architecture includes a provider-agnostic streaming layer; a dictation module for one-way audio; Ramble (our “brain dump” module); and a conversation module to support streaming bi-directional audio and future conversational features.

This layered design means we can ship new voice features with minimal additional infrastructure work. It also enables provider flexibility; although we’re using Gemini Enterprise Agent Platform in production, our abstraction layer also easily supports other solutions.

In addition to helping us tackle three of our four key technical challenges, Agent Platform delivered some nice surprises. First, session resumption was easier than we expected. We initially thought maintaining conversation state across reconnections would require complex server-side session management. But once we understood Agent Platform’s resumption token approach (the token is provided by the API and changes with each context update), implementation was straightforward across all platforms.

Second, context injection worked on the first try. We spent considerable time designing how to provide user context (projects, labels, preferences) to the model. We explored complex retrieval strategies and dynamic context windows. In the end, the simple "v1" approach—just passing most of the user's metadata in the system prompt—worked remarkably well.

For testing, we combined structural validation (task count, priority levels, date presence, etc) with semantic validation (did the model understand the user's intent?) following the LLM-as-judge approach. A second Gemini model evaluates whether the output semantically matches the expected outcome. Native speakers from our global team recorded real-world scenarios in their languages and local accents (15+ language variations and over 100 recordings total), with each scenario having expected semantic outcomes (e.g., "should create 3 tasks: one about calling family, one about shopping, one about exercise on Saturday at 11 AM").

We then created a defined pass-rate threshold for the test suite overall, while also monitoring per-language performance to catch regressions. This approach lets us evaluate new model versions systematically, understanding not just overall performance but also which specific languages might see improved or degraded experiences, and make data-informed decisions.

Ultimately, Ramble is a resounding success in helping our users handle the chaos of day-to-day life. It joins the ranks of Todoist’s Quick Add — our existing natural-language task input — in providing yet another way to capture tasks that is the best in its category.

UKG unlocks real-time workforce intelligence at scale with the Agentic Data Cloud

Wed, 29 Apr 2026 16:00:00 +0000

At UKG, we’ve spent years building and expanding our human capital management (HCM) and workforce management (WFM) solutions with new products, capabilities, and a series of acquisitions. Our cloud platform includes a suite of connected systems that support every corner of the employee experience, including scheduling and workforce operations, HR and payroll, and culture and engagement tools.

These connected tools offer customers incredible depth, but it also means our backend reflects years of evolution. We have 126 application teams, dozens of tech stacks, and more than 12,000 database instances inherited through acquisitions and product growth. And each product carries its own schema and operational footprint.

Previously, data moved through bespoke pipelines not built for real-time use. As AI advanced, expectations did too. Customers wanted instant insights across HR, time, pay, culture, and operations, and those insights increasingly needed to drive automated workflows and intelligent applications.

Internally, teams needed consistent, high-performance access to shared data to innovate faster and modernize our architecture. We needed a unified foundation for the next generation of intelligence across our suite. That’s why we built People Fabric, our new data and intelligence platform powered by AlloyDB for PostgreSQL and the just-announced Agentic Data Cloud.

Unifying the systems behind the suite

People Fabric started with a simple need: bring the full UKG suite onto one real-time foundation. Getting there started with defining a single canonical data model for the entire suite. This would serve as the shared language for people, work, pay, and culture data — consistent no matter where the information originated.

We needed an operational database that could ingest changes quickly and scale horizontally. That’s why we chose AlloyDB as the core of People Fabric. It gives us millisecond-level read-after-write behavior, high-throughput ingestion, scalable read pools, and native vector capabilities to support AI.

With the model defined and the operational store selected, the next step was building the pipeline that feeds the platform. We created a custom change data capture (CDC) framework to extract changes from our existing operational databases inherited over the years. Those changes flow through Dataflow, where they’re transformed into the canonical structure that AlloyDB for PostgreSQL expects.

Once in AlloyDB, that data becomes the real-time backbone of the platform. Applications use it for near-instant queries. AI agents rely on it for cross-domain decisions, and vector search engines use it to power natural-language and similarity-based experience layers.

For larger analytical workloads, the same data flows into BigQuery, which gives our teams and our customers the ability to perform organization-wide reporting and analysis without straining the system. Cloud SQL holds the metadata and tenancy context that govern who can see what and how different parts of the suite interact with People Fabric. From there, the system runs continuously. Data enters through streaming ingestion and gets modeled once in AlloyDB for PostgreSQL to make it available everywhere.

aside_block: <ListValue: [StructValue([('title', 'Build smarter with Google Cloud databases!'), ('body', <wagtail.rich_text.RichText object at 0x7fdc32d05cd0>), ('btn_text', ''), ('href', ''), ('image', None)])]>

Bringing people intelligence to intelligent people

With the architecture in place, People Fabric gives us something we never had before: a complete and consistent view of people, work, pay, and culture data that’s updated continuously and ready for AI to use in real time.

That unified context is what powers our assistive experiences, including conversational reporting and natural-language interactions. Leaders can ask questions in plain English and get answers that reflect the full picture — not just a single system’s slice of it.

With the power of Google’s Agentic Data Cloud, our platform unifies analytical and transactional data to power real-time AI. This allows agents to reason over live workforce signals and trigger immediate actions. Because this data is governed and modeled from the start, our agents can reliably handle multi-step workflows across HR, payroll, and timekeeping.

Whether they're identifying pay discrepancies, adjusting schedules, or flagging compliance risks, they operate with the same shared semantics and security model that guides our applications. It’s the difference between AI that reacts and AI that can truly assist.

Driving impact across every layer

For engineering teams, People Fabric acts as a database-as-a-service that removes the need for each microservice to manage its own datastore or pipelines. This accelerates development and supports modernization without customer disruption.

AlloyDB for PostgreSQL delivers millisecond read-after-write behavior, zero replication lag, and near-real time ingestion latency, enabling real-time workloads with far less complexity. Migrating core person and employment data off our on-prem monolith has generated cost savings significant enough to fund half of People Fabric.

Real-time operational data now gives managers a live view of staffing, pay, and workforce activity. More than 1,000 organizations are already on the platform, with another 1,000 in progress.

As we continue expanding People Fabric, we’re laying the groundwork for deeper agentic automation, more responsive analytics, and a growing set of AI-driven capabilities — all on a trusted, scalable foundation built for what’s next.

Learn more

UKG’s success illustrates how leveraging AlloyDB for PostgreSQL and the Agentic Data Cloud allows organizations to unify operational and analytical data, creating the essential foundation for real-time, agentic AI.

Learn more about AlloyDB for PostgreSQL and get started with a free trial today!
Sign up for a strategy workshop today on how to get your data ready for the agentic era!