Saket S.

I've been following the context graphs debate with great interest. Jaya Gupta and Ashu Garg's "trillion-dollar" thesis, Prukalpa ⚡'s perspective on heterogeneity, Arvind Jain connecting it to transactive memory systems. These are important ideas. And the problem they're highlighting is real. Enterprises desperately need better ways to capture how decisions get made. Having spent years trying to build exactly these systems, I want to share what I've learned about why this is so hard. Lesson 1: You Need to Own the Interface At Nutanix, I worked on Intentful APIs for cloud infrastructure. The idea was to capture not just what changed, but why. We learned something fundamental: this only works if you own the entire interface. Kubernetes can do this because everything flows through its control plane. But enterprise environments are messier. States get mutated directly in underlying layers. Someone SSHs into a box. A quick database fix happens at 2am before a board meeting. This isn't bad behavior. It's reality. But it means any context graph is working with incomplete information. Lesson 2: Decisions Bottom Out at Human Judgment At Kognitos, I built exception handling systems which captured "decision traces". Decision-making is deeply recursive. Ask why a discount was approved. Then why that exception exists. Keep asking why. Eventually, you arrive at: "Because someone with experience made a judgment call." That's not a failure of documentation. That's how organizations work. Decisions happen under time pressure, with incomplete information, based on a team's evolving understanding. Lesson 3: The "How" is Often Private Recently at SoraNovaAI, I built knowledge base systems for a venture capital firm. What I discovered is that people often keep the "how" of their decision-making private. Not out of malice, but because it's genuinely their edge. A senior investor's pattern recognition. A sales leader's intuition about which deals to push. These are competitive advantages tied to personal goals. Context graphs assume everyone wants their decision-making captured. The reality is more nuanced. I believe in the vision of better decision capture. The pain is real. Organizations lose institutional knowledge constantly, decisions get relitigated, context disappears when people leave. But different people always see different parts of the same decision. That's not a bug. It's how organizations work. What I've learned: - Design for incomplete context, not perfect capture - Focus on domains where incentives align with transparency - Accept that the last mile is human judgment - Celebrate progress over perfection The real world is messy. The opportunity is in building systems that embrace the mess and still make things better. I'd love to hear from others who've navigated this. What's worked? What hasn't? #contextgraphs #enterpriseai #aiagents

59

17 Comments

Ex-Meta manager says Stanford and MIT CS grads struggling to get jobs due to AI, shares top skills to get hired - India Today https://lnkd.in/gJDZqGw5 #artificial #intelligence #digital #transformation We are living through another revolution. This time, it's down to artificial intelligence. It is already changing the society we live in and is expected to reshape how we work, do business and interact with technology, eventually transforming both the economy and wider society. But right now, the most immediate impact led by AI is being felt in the job market, even bringing challenges for the highly educated. In fact, according to a former Meta executive, students from elite institutions such as Stanford and MIT are currently struggling to secure job roles, as companies are rethinking hiring strategies.

1

Proud to help power conversational AI at scale for JioHotstar — transforming content discovery into real-time, personalized engagement for millions across India. https://lnkd.in/gjnaZDmA

4

Ruokai Yin, our summer intern from Yale to Microsoft Azure, worked with Xuan Zuo, @Sattwik Deb Mishra, Hokchhay T. and Preyas Shah and me. We published a paper based on his project: https://lnkd.in/eHuG3BSt Summary of paper: Finding optimal sharding is hard especially for large LLMs. A large number of devices have to be divided into TP, PP, EP dimensions. For TP, we have to decide the tensor dimension e.g. head, hidden size to shard for every op. Additionally, TP can be one-dimensional/two-dimensional etc. This makes for a large sharding space that is hard to cover exhaustively. SoTA methods use heuristics to select sharding e.g. GSPMD, Sequence Parallelism, Helix etc. There are significant time gaps, months or years, between the discovery of each of these methods. We want to fast-forward the process of discovery by 1) using a formal representation of the sharding space, and, 2) using reinforcement learning to search the sharding space.    This paper demonstrates that our method is 3.5x better than random search, and on par with hand-tuned heuristics when looking only at 1D sharding and without allowing sequence parallelism. We aim to extend the work to multi-dimensional sharding as well as sequence parallelism. Beyond that, we wish to combine the problem of sharding with fusion, tiling and collectives as well.

58

Agentic AI role is exponentially increasing in software development lifecycle. Janus is an agentic tool developed by Vayavya Labs that automates part of this lifecycle for code compliance. For more information meet us at nasscom ai

49

4 Comments

🌟 From curiosity to stage: my journey with DSPy at The Fifth Elephant A week ago, I had the incredible opportunity to speak at Hasgeek’s Fifth Elephant Conference — one of India’s most vibrant gatherings for data and AI enthusiasts. When I first started exploring DSPy (Community), I was fascinated by how it simplifies building structured, reliable AI systems — turning messy prompt engineering into a declarative, modular process. Over time, that curiosity turned into something bigger — a belief that DSPy represents the next step in how we design AI systems. In my talk, I shared how this framework can help us: 🔹 Build reasoning-centered LLM pipelines 🔹 Reduce prompt fragility through programmatic supervision 🔹 Apply these ideas to real-world problems 🎥 If you missed the session, the full recording is now available — and I would love for you to check it out: 👉 https://lnkd.in/dRBSFRF5 Whether you are building AI agents, working with LLMs, or just exploring what’s next in applied AI — I hope this talk sparks a few new ideas. #DSPy #AI #LLM #MachineLearning #fifthel #Hasgeek #AIEngineering #LLMAgents #ShiftScheduling #OpenSource Hasgeek

73