One of the things I love most about my job is hiring great talent. Being in interviews, learning about people's journeys, figuring out how they fit into the team and what unique strengths they bring - it's genuinely energizing. And every once in a while, you meet someone from the top 1%. Those conversations are different. You stop interviewing and start brainstorming how you'd build great products together. We're hiring for multiple product roles at Sarvam. If this sounds like you, come talk to us: sarvam.ai/careers

Another day, another best in class model from @SarvamAI Huge credit to the TTS team for this one. Built for voice agents and content creation, and it really shows. Try it: indus.sarvam.ai

One more killer product from Sarvam is here. Hands down the best dictation app. Give it a try. heykivi.ai/

Agastya turns 1 today ❤️ Easily the most fulfilling 365 days for me. We are beings of extraordinary potential and it takes circumstances for us to realise it. Having a child made me realise how much love there's in each of us to give, how much patience and endurance we are capable of and how rewarding life can be.

Got this email on our support platform yesterday. A small trader in Madurai used our AI to study 5 months of his own statements. It showed him where he was losing money and helped him set rules: 1% risk per trade, stop-loss before entry, out by 3:05. His words: "For the first time, my mind is completely at peace during market hours." This is why we build. 🇮🇳

Jev giving up text generation on purpose is beautiful. Software was always structured: databases, tables, API contracts. Then LLMs came along and made the interface text. We got structured output back, but only by burning more tokens and compute and still writing “please adhere to the JSON schema” to frontier models. A few things clicked for me. Look at an agent trace. The first chunk is usually the model deciding which tool to call. That’s not really generation. It’s a decision. Give that to Jev. Let it pick the tool, execute it, and bring the LLM back only when there’s something worth saying. Same for routing and classifier models you use at the start of agentic turns. RLCD models like jev handle the decisions; expensive models handle the reasoning and generation. The second implication is the wild UX possibilities that get unlocked because of low latency. Imagine a feed where a model classifies every post against what you actually care about and filters the rest before it ever reaches you. I’m hoping the next class of models solve of open ended intent and entity extraction. Wild wild times.

We are over-compressing a whole product org into the title “AI PM.” One thing I’ve seen clearly: once you operate agents at meaningful scale, the work starts decomposing into distinct product disciplines. 1. Agent Platform / Harness PM Owns the underlying agent system: core loop, sub-agents, context, memory, compaction, skills, sandboxes, tool execution, permissions, model routing. These are deeply technical choices. In practice, I’ve found that PMs working here need enough technical depth to make them well. 2. Evals / Observability PM Defines what “good” looks like, how failures are measured, and how production behavior feeds back into improvement. This also feels like one of the least invested areas across companies today. Because the space is still nascent, the eval stack itself becomes a product decision: Braintrust vs. Langfuse vs. build your own; what to log, how to inspect traces, replay failures, and connect offline evals to production. At scale, I increasingly think evals should drive a large part of the agent roadmap. 3. Runtime / Experience PM Owns how the agent behaves and how that behavior reaches the user. When should it search, think longer, ask a question, make an assumption, or use a tool? And what should an answer be: text, table, chart, dynamic card, artifact, generated UI, voice, vision? A lot of our product work has made me believe we are still very early in discovering the native UX of an agent. There are other layers too: frontend/SDKs, streaming, developer experience, etc. But the pattern feels familiar. As product management matured, it split into consumer, growth, platform, gaming, infra, and other specialties as each area developed enough depth to become its own craft. The same thing will happen to AI PM. “AI PM” is just a temporary abstraction.

A prompt is all input. You describe what the agent knows, what to consider, what to be careful about. There is no line in a prompt where you say what a right answer looks like, so you leave that to the model. The model then decides, on its own, how long the answer runs, whether it asks or assumes, how much of its work it shows, what it does when it is half sure. It decides differently every time, and differently again when you change models. Writing the exchange first forces those decisions out of the model. You cannot write a transcript without being clear on what great answer looks like. Quality went up sharply for us when we started this way. Write what a right response looks like before you write the prompt that is supposed to produce it.

You need 20 people to go from 0 to 80. Then you need 80 more to go from 80 to 100.

Happy Teachers Day to everyone who taught me something this year. Also to the LLMs, who answered the same question four different ways at 1am without once making me feel dumb for asking.

Super psyched about this. Newer models dropping soon. Stay tuned.

What a beautiful cover image. Says the whole thing without saying anything.

Great managers are exceptionally good ICs.