Chirag Dadia

💥 That moment when you run a “harmless” ALTER TABLE in production… and your DB screams louder than your morning coffee ☕ What if I told you your database can actually time travel? 🕰️ In my latest article, I show how Liquibase turns “oops” into “no worries” with version-controlled migrations, rollbacks, and traceable schema changes. Let me know what you think!

8

Agentic AI workloads are fundamentally different. Unlike chat or summarization, agentic workloads accumulate context across many steps, often reaching hundreds of thousands of tokens. That makes measuring performance on real-world agentic sessions essential. The first on-silicon Vera Rubin NVL72 results, measured by NVIDIA on the SemiAnalysis AgentX workload and compared with GB300 NVL72 running DeepSeek V4 Pro, show: • Up to 30x higher throughput per megawatt • Up to 35x lower cost per token Each generation of hardware and software improves the performance and economics of NVIDIA infrastructure, helping us scale increasingly demanding AI workloads. #AI #AgenticAI #NVIDIA #VeraRubin #Blackwell

92

2 Comments

Modern data strategy isn’t one-size-fits-all. Different architectures serve different purposes, and many organizations use a combination to balance scalability, governance, and analytics needs.

12

What I find interesting about the RLM (Recursive Language Model) approach is its potential to be a standard solution for the long-context problem — much like what we’ve seen with CoT, ReAct etc for reasoning Like CoT and ReAct, RLM builds on a recursive framework that uses the model itself to iterate — but here, the focus is on tackling the long-context problem. The paper is co-authored by Omar Khattab, the creator of DSPy, so it might just be a matter of time before we see RLM technique integrated into DSPy (Community). https://lnkd.in/gTy3ghxe

10

2 Comments

Cassandra (Cassie) Shum, VP of Ecosystem Product Engineering at RelationalAI, on whether LLMs have made knowledge graphs redundant: The retrieval argument for graphs was stronger six months ago. Models have caught up on that dimension. The more durable argument is treating the knowledge graph as the foundation of your system rather than as a query layer. More sessions like this are lined up for QCon AI New York 2026 (Dec 15-16). Find out more about QCon AI New York here https://bit.ly/4yDXymz

1

Call to action for AI Providers developing platforms and tools in the Global Coaching Profession 👇 From Chatbots to Coaching Partners: Building Trustworthy Agentic AI Bani’s post nails a core truth about the evolution from RAG to agentic workflows: If you can’t measure the intermediate steps, you’re not building an Agent, you’re building a slot machine. In the coaching profession, that truth carries special weight. When AI enters a relational, reflective, human space like coaching, we’re not just optimizing reasoning loops or API correctness, we’re shaping how people experience technology and how the technology enables or undermines trust, empathy, and growth. For organizations developing AI for the global coaching profession, the evaluation frontier extends beyond logic and tool use. We must also unit test for: Ethical Trajectory: Does the agent’s reasoning align with coaching values and standards like facilitate client insight, inclusion, and client autonomy? Psychological Safety: Does the interaction maintain a sense of client self-expression, or does it slip into control, advice, or bias? Learning Fidelity: Is the agent’s adaptation improving reflective capacity, or simply optimizing response efficiency? These dimensions demand “glass box” visibility — not just into how agents think, but into how they relate. Agentic AI isn’t a coach; it's a tool, an enabler that supports the coaching process. Coaching is a human discipline that tranforms lives. That means building evaluation pipelines that combine Bani’s technical rigor (DeepEval, RAGAS, CI/CD metrics) with human-centered assurance frameworks that measure dignity, trust, and growth outcomes. So I’m curious: - What are you measuring when you test AI built to enable human development, learning, or coaching? - How might we align evaluation standards across engineering and ethics to make “production-grade” also mean “people-safe”? Let’s make sure that as our agents get smarter, our systems get more humane. International Coaching Federation #AgenticAI #AIforCoaching #AIArchitecture #EthicalAI #LLMOps #DeepEval #RAGAS #HumanCenteredDesign #TrustandSafety #CoachingInnovation #ICF30

11

1 Comment

From Prototype to Production: Evolving Vector Search with LangGraph & Azure When building RAG workflows, starting with a heavy managed search service right out of the gate can add unnecessary complexity. Instead, stepping through a progressive architecture helps ensure every piece earns its place. In this 3-part series, I break down how to build and migrate a vector search pipeline step-by-step without rewriting core application logic: 1️⃣ Part 1: Ephemeral In-Memory Vector Search Starting with the simplest possible approach—using LangGraph, Azure Document Intelligence for OCR/markdown extraction, Azure AI Foundry, and an in-memory vector store for session-isolated, single-conversation chat. https://lnkd.in/ev59t87p 2️⃣ Part 2: Moving to Redis for Persistence & Scaling Swapping in-memory storage for Redis Vector Store. Introducing auto-eviction via TTL, index-per-session management, and cross-request persistence while analyzing real memory footprint overhead. https://lnkd.in/eQGXXFYg 3️⃣ Part 3: Enterprise Scale with Azure AI Search Migrating to Azure AI Search using its push API. Moving our custom ingestion/chunking pipeline directly into a shared enterprise index with hybrid search capabilities, security trimming, and production-grade scale. https://lnkd.in/ek3WkEgc Key Takeaway: By modularizing the retriever node in LangGraph, swapping storage providers from memory ➔ Redis ➔ Azure AI Search requires changing minimal code while unlocking production capabilities. #AI #LangChain #LangGraph #Azure #Redis #VectorSearch #RAG #CloudArchitecture #Python

39

4 Comments

We used to build NLP pipelines that looked like a jungle (tokenizers, embeddings, intent classifiers, feature stores). Then came LLMs, promising to replace all that with a single prompt. And for a while, it really did feel magical. Until teams started realizing that “one model fits all” often means “one model fits no one.” Because in the real world, NLP isn’t just about fluent text. It’s about understanding your domain, your jargon, your exceptions, your context. Here are my two cents: - If you treat “use an LLM” as a shortcut, you’re probably creating a new form of debt. - But if you treat it as one part of your system, combined with feature engineering, domain lexicons, and context pipelines, you’re building something reliable, something that lasts. At the end of the day, NLP isn’t about bigger models. It’s about right-sized models, tuned to your world. #RightComplexity #AIWithoutMyths #EngineeringReality #NLP

26

As AI agents move from experimentation into production, we need to rethink how we operate and monitor them. I recently wrote for APM Digest about a simple question:  If an AI agent becomes a production dependency, should we manage it like one? The article explores practical challenges around observability, reasoning traceability, hallucination detection, governance, multi agent workflows, and cost. Sharing it in case it helps others thinking through what production ready agentic AI should actually look like. The article link: https://lnkd.in/dG77rPhS #AgenticAI #AIObservability #AIOps #AIEngineering #MLOps

5