Product engineer. I build products that use AI agents, and the parts that keep them trustworthy once they act on their own: retrieval, tool use, evals and failure investigation.
Website · Writing · LinkedIn · X
- LLM products: RAG, tool-using agents, citations and checks for domains where a fluent wrong answer is a failure
- Evaluation: benchmarks, synthetic tests and tracing a bad output back to its cause
- Full-stack delivery: from talking to users through design, build and production
- Systems that touch the real world: queues, retries and firmware, where a failure has to end in a defined outcome
- Graphify benchmark: a 360-trial test of whether a code graph helps coding agents answer questions. Plain search was hard to beat, and I published that.
- Graphify contributions: seven fixes released upstream.
- Ellie: my WhatsApp assistant. Synthetic tests run every night and a coding agent proposes fixes that I review. Example PR.
- undupe: finds duplicate open pull requests on busy GitHub repos.
- Pinsight: turns public GitHub threads into source-linked product signal.
- Crediflow AI: production LLM workflows for lending, built with the analysts who use them
- Wendor: reusable frontend systems, and software connecting payments, cloud queues and physical machines
- Klarity (YC S18): retrieval and evaluation on legal documents
Stack: Python, TypeScript, Go · React, Next.js, FastAPI, Node.js · LangChain, LlamaIndex, PostgreSQL/pgvector · AWS SQS, Inngest, Docker



