Anuradha Bajpai

From the front line of delivering Agentic systems Forget the pilot and the launch party. The real game of agentic AI starts when it goes live. In 2026, a lot of Agentic systems are shifting from isolated pilot programs to full enterprise-scale deployment. Being in the middle of this shift, I see these areas to focus on 1. Token Discipline. This is the key to control cost for Agentic systems. Treating tokens like a precious currency and implementing real-time metrics for every agent request. Without this, costs spiral and the ROI vanishes. Tokens are the measure of Cost Of Goods Sold (for those of you in finance LOL). 2. Think agent Observatory, not dashboards. Develop or integrate with systems that evaluate agent accuracy, detect subtle failures, and trigger automated playbooks is top of mind. I see observability as the smoke detector for customer trust. Without one we are flying blind, and the first one who will notice that our system doesn’t deliver value is…our customer. 3. Institutionalize Lineage. This often overlooked. Every data source, prompt version, model iteration, reasoning result and evidence should be traceable. Version everything. And in all this, my observation is also that the definition of the secret sauce, our intellectual property, and competitive advantage, has shifted from traditional AI models a few years ago to the agentic reasoning prompts and the data layers and knowledge graphs powering these agents. Interesting times! #AIStrategy #AgenticAI #LiveAgents 2026

23

Customers are asking for AI that goes beyond assistance to true collaboration. Claude Opus 4.5 from Anthropic delivers on that vision: understanding objectives, factoring in constraints, and ready to power complex workflows. Opus 4.5 joins the full set of Claude models now in Microsoft Foundry, alongside models from OpenAI, Mistral AI, Cohere, Meta, and others, so you can easily compare and find the best-fit models for your needs. With Microsoft Foundry, Azure customers get the broadest selection of model providers on any cloud, plus the governance, security, and interoperability to accelerate innovation with confidence. 🔗 https://lnkd.in/gdRxNB_q

106

3 Comments

AI acceleration > AI bubble narrative after NVDA’s Q3 beat. From the transcript -- “demand for AI infrastructure continues to exceed our expectations. The clouds are sold out and our GPU installed base, both new and previous generations, including Blackwell, Hopper, and Ampere, is fully utilized. Record Q3 data center revenue of $51 billion increased 66% year-over-year, a significant feat at our scale. Compute grew 56% year-over-year, driven primarily by the GB300 ramp, while networking more than doubled given the onset of NVLink scale up and robust double-digit growth across Spectrum X Ethernet and Quantum X InfiniBand.” GSXAPOWG (GS Asia Electrification) sharply recovering today, seeing the second best day YTD. That said, macro picture turned incrementally “negative” overnight after hawkish Fed’s Oct minutes. Dec cut probability lowered to 36% from 50% overnight. USD/JPY at 157. VIX still above 20. Still not out of the woods .. 

1

NVIDIA and Palantir Technologies have teamed up to bring sovereign AI into critical supply chains, starting with NVIDIA's own operations. The stack pairs NVIDIA Nemotron open models with Palantir Foundry and AIP, built on the Palantir Ontology, giving teams shared visibility into supply constraints. Alex Karp, cofounder and CEO of Palantir, called it "the most valuable, intricate and complex supply chain in the world." Jensen Huang, founder and CEO of NVIDIA, said the firms are "transforming this vast operational graph into sovereign intelligence." Other organisations can build similar systems on it too. #NVIDIA #Palantir #SovereignAI #SupplyChain #ArtificialIntelligence

172

1 Comment

An LLM doesn't write a sentence. It predicts one token, then another, then another — thousands of times per response. Two completely different phases hide inside that loop: prefill (reading your prompt, compute-bound) and decode (generating the reply, memory-bandwidth-bound). Nearly every "speed up your LLM" technique you've heard of — KV caching, prompt/prefix caching, semantic caching, quantization — only makes sense once you know which of these two phases it's actually fixing. Mix them up, and you'll misconfigure your caching strategy, overpay on inference, or chase the wrong bottleneck entirely. In this piece I break down all three caching layers — what each one actually does, how long it lives, when to use it, and the #1 mistake that silently kills your cache hit rate. If you're building with LLMs in production and have ever wondered why your "cached" prompt isn't actually getting cached, this is for you. 👇 Full breakdown in the article https://lnkd.in/gTD6iAJN #LLM #AIEngineering #LLMOps #GenerativeAI #AIInfrastructure

9

1 Comment

Great time at the Liquidity Summit two weeks ago. I had the chance to sit down with the NYSE right after Bill Ackman to talk AI investment trends, compute demand, and what's driving the next phase of growth. Here's the conversation.

93

1 Comment

The "AI Feature Value" conjecture for B2B SaaS (Adapted from https://lnkd.in/eCxQebT7) An AI feature or an Agentic AI workflow only creates value within a product based on two properties: Data Locality and Data Assembly. That is, there are tons of valuable uses for LLMs, but the continued growth and improvement of open-weight models and self-hosting means that most AI features will be runnable locally or within private networks in the future. Thus, for a product to charge money for an AI feature specifically, the product must possess a rich and unique data locality, and/or the AI feature must comprise a rich and unique data assembly. **Data Locality** If the product contains a unique set of (customer) data, then there is value in using an LLM to interact with that data. This is analogous to paying a premium for a bottle of water at the airport or the stadium. The provider has created a valuable environment that the customer can't (or doesn't want to) create and maintain on their own, and the customer pays for additional enhancements to their experience there. **Data Assembly** If the AI feature takes data sources and transforms or distills them in a non-trivial way and combines them with external sources or knowledge and manipulates the result in an iterative or compounding way, then the result has value and, by extension, so does the data construction, or assembly. The key to such an AI feature is that it decomposes the problem in some arbitrary domain into fundamental patterns and relationships of information, which is the language learned and spoken by LLMs. Analogously, a client brings all of the facts (theoretically 😄 ) to their attorney who organizes the data into a coherent argument. The lawyer gets paid to take raw data and translate it and decompose it into logical units that align with the universal truths of the domain. Of course, the risk to making money on Data Assembly is that the world might eventually figure out the decomposition and then everyone can replicate it locally for use with their own LLMs. This suggests that the most defensible problems to attack with LLMs are the most complex ones. **Summary** The best way to judge an AI feature or workflow is to ask: What is the customer paying for? It must be either: - The rich data environment that will be exposed to the LLM (good!), or - The proprietary data preparation and assembly performed before presenting it to the LLM (great!). If it's both, then that's the golden egg-laying goose (🤑). If it's neither, then it's back to the drawing board (😔).

24

People went ballistic on X when MIT’s State of AI Report claimed 95% of enterprise AI projects fail to deliver ROI. The stat isn't false, but the real story is more nuanced... Inside enterprises, two things are happening: First, orgs try to build AI tools in-house. Their internal IT or data teams might be great at coding, but they’re rarely set up to deliver tools tailored to the specific workflows of each department. Projects stall in “pilot” because they don't fit how people work. On top of that, internal departments often operate in silos, making it nearly impossible to build AI tools that work end-to-end across the org. The result is a patchwork of PoCs that don’t scale. Second, when DIY falls short, enterprises turn to consultants. Department heads compile their wish lists, and suits deliver an impressive-looking strategy deck. But implementation usually gets thrown back over the fence to internal teams - who are already struggling - which means many companies end up right back where they started. Reading this as “AI is overhyped” is a false oversimplification. What’s really happening is a gap in how orgs use AI. That gap is a wedge for startups. After burning years and millions on internal experiments, F500s are increasingly looking for founders who deeply understand their use cases and can ship AI solutions that plug into their workflows. Since most enterprises struggle to build and deploy these systems themselves at scale, they’re more willing than ever to bet on startups who can. While the echo chamber on X keeps yelling that “AI is useless,” sharp founders are snagging the contracts legacy players keep whiffing on.

176

47 Comments

💥Mentorship and collaboration opportunity on MidcurveLLM is now open. 💥 ➡️ This project focuses on fine-tuning LLMs for geometric midcurve generation. ➡️ You will work on non-NLP, geometry-centric datasets using text-based formats. ➡️ The goal is to generate midcurves from 2D closed shapes with high precision. ➡️ Learn why fine-tuning works better than RAG for geometry problems. ➡️ Get hands-on with LLM fine-tuning, transformers, and benchmarking. ➡️ Ideal for practitioners with LLM fine-tuning and Python experience. ➡️ This is a voluntary, research-focused learning opportunity. Read the full post here: https://lnkd.in/dbXNgYfu #spatialai #AI #LLM #FineTuning #GeometryAI #Research #OpenSource

29