Worked on a hybrid retrieval stack that combined semantic search, lexical matching, and a re-ranking layer on top, tuned for the relevance and consistency requirements of LLM-powered applications. Built the evaluation pipeline that scored the system on precision, recall, and ranking metrics, against LongMemEval and LoCoMo (the two canonical long-context memory benchmarks) and against competitor systems. Those results became the signal the team used to decide what to ship next.

Also built the internal admin tooling, frontend and backend, that lets the team inspect, debug, and analyze retrieval behavior on real traffic. That tooling was the difference between iterating on intuition and iterating on evidence. Contributed to the system-level calls on identity consistency, embedding model selection, and re-ranking strategy.

Built the product end to end as the only engineer. A Python and FastAPI backend with PostgreSQL and Redis, LangGraph agents orchestrating outbound research and outreach, async workers, real-time dashboards over Socket.IO, human-in-the-loop review for every generated message, and a Next.js app in TypeScript and Tailwind CSS on top, with React Query for data and Auth0 for auth. Owned the infrastructure, the prompts, the evaluation, and the design.

One piece worth singling out is the composable prompt system. Each campaign injected the performance signal of previous campaigns back into the prompt, so generation quality drifted upward over time without anyone touching a template.

The most useful thing I learned in that year was how fast a LangGraph stack accumulates cost and reliability debt when nobody is watching. Most of my current writing is downstream of that observation.

Ran five engineering teams, thirty-five plus engineers reporting through me. The core piece of work was rebuilding the entire payment workflows and pricing models end to end, which brought processing errors down by 40%.

Around this time I started the Berkeley ML/AI postgrad, which is what eventually pulled me back into hands-on building.

Joined as a senior engineer, ended as a manager running a product-area team. Useful stretch, because I got to do the transition inside a single company.