I build production AI agents and the evals that check them, and I'm learning to train agents with RL.
Agents in production
As AI Tech Lead at Coefficient I lead a team building a multi-step agent platform: multi-agent orchestration on LangGraph and Deep Agents, persistent memory, retrieval (RAG) and enterprise search, natural-language SQL, and scheduled, long-running workflows across 100+ enterprise connectors. One system I shipped runs six cooperating subagents for monitoring and alerts.
I built what agents need to do real work safely: sandboxed Python execution, authenticated browser sessions and workspace-scoped data access on autoscaled infrastructure. And I care most about knowing whether an agent actually did the job: 100+ agent evals that combine deterministic graders with LLM-as-a-judge scoring, traced failures, and eval-driven model routing, prompt caching and context compaction that cut LLM inference costs by 60% at comparable quality.
Before
Eight years of backend and distributed systems at Coinbase and Flipkart.
Research
An explicit family of counterexamples to the Jacobian conjecture in every dimension n ≥ 3, in two preprints, later discussed on Terence Tao's blog. Now: post-training, RL for agents and tools, verifiers and evals, learned by building and posted every Sunday as a lab log.