I build the knowledge layer AI agents work from.

Founding engineer at Pavo AI. Before that, seven years scaling recommendation at ShareChat from one million users to two hundred million, then code search at Sourcegraph. I write about what agents lose between finding something and using it, and I design interfaces on both sides of an agent: the ones people use to direct and check it, and the ones it works through.

Papers at WWW, RecSys, SIGIR and CIKM Keynote at FIRE 2023 Open source: crewlore

What I believe about building with AI

  1. Evidence is only useful if it survives every handoff. 1 Retrieval Is Not Delivery Of the excerpts one agent had read, 22% reached the next agent intact. A structured record, with the changes around it, took that to 60%.
  2. An agent is finished when it can prove it, not when it says so. 2 Completion Is a Proof, Not a Status A tool call can end while the command it started is still running. Four clocks have to settle first.
  3. Measure how much a benchmark moves before claiming it improved. 3 Your Eval Has a Noise Floor One unchanged configuration scored 33, 27, 17 and 25 claims out of sixty.
  4. Choosing what an agent sees is a ranking problem. 4 Agent Context Is a Recommendation Problem Picking a small, ordered set under a hard budget when nobody asked is a close cousin of what recommenders have worked on for twenty years.
  5. An agent forgets things it is still holding, so write them down when it sees them. 5 Context Aging An agent read a two-step rule in its first minute and drafted it, more than twenty minutes later, as one step.
  6. A person’s correction should outrank the agent that wrote the draft. 6 Reviewing what an agent wrote In the review prototype, the agent can flag a sentence someone corrected and propose a change, but only a person can make it.

Writing

All writing
Context Aging: What an Agent Forgets While It Is Still Running Sep 2026 · 17 min · with Susnato Dhar An agent can lose a fact it is still holding: read in its first minute, never truncated, missing from what it writes twenty minutes later. A working name for that, what we built against it, and a null result. Agent Context Is a Recommendation Problem Sep 2026 · 12 min Deciding what an agent sees, out of everything it could see, under a hard budget, with position effects and almost no feedback. Recommender systems have spent twenty years on a close cousin of this problem. Your Eval Has a Noise Floor Sep 2026 · 12 min Before you can tell whether a change helped, you have to know how much your benchmark moves when nothing changes at all. Usually nobody has measured that. Completion Is a Proof, Not a Status Aug 2026 · 10 min · with Vivek Sharma A coding agent is not finished when its stream closes. It is finished when its processes are dead, its output is drained, and its workspace is durable, and the runtime can show evidence for all three.

Design

All design
Reviewing what an agent wrote AI-human interface · Oct 2026 Four places a review of an agent's document goes wrong, and a prototype with an answer for each.
When knowledge reaches the agent AI-native interface · Oct 2026 One team rule and a ten-turn coding session: where each delivery route leaves the rule when the agent edits.

Open source

GitHub
  • crewlore

    A local-first compiler that turns coding-agent sessions and agent-authored pull requests into a plaintext, citable team-knowledge book in your git repo. Every claim quotes the exact source line it came from, and records whether the team adopted it or tried it and declined. Alpha, on PyPI: pipx install crewlore.

  • agent-relay

    Multi-agent workflows that learn from their own past runs. Plans, reviews, build logs and audits live in your git repo as markdown, and relay distill --llm turns past reviewer rejections into lessons the next planner reads. Alpha.

Papers

Google Scholar

Speaking

All talks

Career

  • 2025–now Pavo AI, founding engineerThe knowledge layer AI agents work from; building the founding engineering team across the US, London and India.
  • 2024–2025 Sourcegraph, machine learning engineerLLM-driven ranking and retrieval for code search over large codebases.
  • 2017–2024 ShareChat, staff machine learning engineerScaled the recommender system from one million users to two hundred million; led a team of 18 to 20 engineers across ranking, moderation and feature pipelines.
  • 2016–2017 Fidelity Investments, data scientist
  • 2011–2016 IIT Kanpur, BS and MS in Mathematics and Scientific Computing

Get in touch

Email srijanskt@gmail.com, or message me on X or LinkedIn. I'm glad to talk about agents, retrieval, evaluation, and interfaces for AI.