Natapon P.

I build AI systems that ship — and I prove they work before you touch them. Most of my code is now produced by coding agents (Claude Code) under my direction. That makes the spec, the integrations and the verification the real work. I don't trust a green CI run or an agent reporting "done": I re-derive the number from the query, check the job actually ran, and confirm the deploy is serving new code. What I do: • LLM agent systems — orchestration, tool use, MCP servers, Claude Code plugins • Evaluation and observability — golden datasets, trace analysis, catching regressions before users do • RAG and knowledge graphs — retrieval with formal reasoning, Postgres/pgvector at production scale • Full-stack delivery — Python (FastAPI, Django), Postgres, React, deployed and handed over • Computer vision — 15+ years, detection, tracking, 3D reconstruction, production pipelines Recent work: a trace-analysis platform in beta where I own the golden-dataset side and the open-source Claude Code plugin that drives it; a booking platform built end to end by agent for a travel client; a stock-screening app in daily use by an investing group. Background: PhD from the University of Tokyo, ~32 peer-reviewed publications, former Assistant Professor. 42 projects on Upwork, 11,000 hours, 100% job success. I work async from written specs, and I hand over in plain English that a non-developer can follow. If you want someone who tells you what's actually true about your system rather than what you hope is true, we'll get along.