Context Aging: What an Agent Forgets While It Is Still Running
Long-running agents don't fail when the window fills. They fail earlier — when what they observed forty calls ago is still in the context and no longer effective.
Building the tribal knowledge layer for LLM agents.
The context engine agents need to reason reliably over enterprise data — with principles that extend to personal-scale context too. Bringing the recommender-systems playbook to a problem the agent-memory world has not recognized as a recommendation problem.
I write here about agents, retrieval, and evaluation. If you are working on the same problems, get in touch.
Selected writing
Essays on retrieval, agent systems, and evaluation. I also co-write the infra + AI newsletter Bits & Neurons with Vivek Chandela.
Agent Context Is a Recommendation Problem
Deciding what an agent sees, out of everything it could see, under a hard budget, with position effects and no explicit feedback. Recommender systems have been solving exactly this for twenty years.
Your Eval Has a Noise Floor
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
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.
Retrieval Is Not Delivery
An agent can find the right evidence and still lose the detail that matters before the final answer. Measuring the whole path instead of the last step.
Crash-Proof Custom Agents
Shipping long-running agents with retries, stops, and durable state.
New essays land in the RSS feed. I post short summaries on LinkedIn and X.
What I'm working on
I'm a founding engineer at Pavo AI in Seattle, building the tribal knowledge layer for LLM agents. My day-to-day is a mix of staged agentic composition, conflict-as-knowledge resolution, and bridging the recommender-systems playbook to agent context. The essays above are the public half of that work — two series so far, on evidence-carrying systems and on context as a ranking problem — alongside research submissions to 2026 conferences and workshops.
Before Pavo, seven years at ShareChat scaling the recommender system from 1M to 200M users, then LLM-driven ranking and retrieval for code search at Sourcegraph (IC-5), and an undergrad & master's at IIT Kanpur in Mathematics and Scientific Computing.
I've built global engineering teams twice — at ShareChat and now the founding team at Pavo, spanning the US, London, and India — and have run 200+ hiring interviews in the process.
Projects
OSS projects and study forks. Everything on GitHub.
- crewlore flagship
A local-first compiler that turns AI-coding-agent (Claude Code) sessions into a
versioned, plaintext, citable team-knowledge layer in your git repo. The compiler
refuses to emit a claim it cannot anchor: every claim ships with a verbatim source
anchor that resolves back to the session line it came from. Compiled end-to-end on
the public pydantic-ai repo (18 claims). 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;
relay distill --llmcompresses past reviewer rejections into forward-looking lessons the next planner reads automatically. Alpha (v0.2). - claude-ui Web interface for Claude with document analysis and conversation history.
- LLMs-from-scratch Study fork — building GPT-style models in PyTorch, from the ground up.
Publications
Published at WWW, RecSys, SIGIR, and CIKM. Full list on Google Scholar.
- Dimension Mask Layer: Optimizing Embedding Efficiency for Scalable ID-based Models. link
- Crafting Tomorrow: The Influence of Design Choices on Fresh Content in Social Media Recommendation. arXiv PDF
- Real-time Event Joining in Practice With Kafka and Flink. arXiv PDF
- Monitoring the Evolution of Behavioural Embeddings in Social Media Recommendation. link
- Formulating Video Watch Success Signals for Recommendations on Short-Video Platforms. link
- MEMER — Multimodal Encoder for Multi-signal Early-stage Recommendations. link
Patents
- PatentComputer systems and methods for representatives to monitor and intervene in robot conversation. link
Book
- BookData Science: Uncovering the Reality. Applied-data-science primer, aimed at practitioners entering the field.
Speaking
Keynotes, paper presentations, and conference service. Most entries link to a page with the slide deck embedded.
- Talk
- Service
- Talk
- Keynote
- Talk
- Talk
Career
- 2025 — Pavo AI·Founding Engineer Tribal knowledge layer for LLM agents; building the founding engineering team across the US, London, and India. Seattle, WA.
- Nov 2024 — Mar 2025 Sourcegraph·IC-5 MLE LLM-driven ranking and retrieval for code search over large codebases — agentic search over code; shipped relevance and quality changes to external users.
- 2017 — 2024 ShareChat·Staff MLE Scaled the recommender system from 1M to 200M users; built and led a global team of 18–20 engineers across 4–5 charters spanning content moderation, ranking, and feature pipelines.
- 2016 — 2017 Fidelity Investments·Data Scientist
- 2011 — 2016 IIT Kanpur·BS & MS, Mathematics and Scientific Computing
Get in touch
The fastest route is srijanskt@gmail.com. I read every DM on X and LinkedIn, and I'm happy to talk about applied research, retrieval and evaluation for agents, and production ML.