Production AI systems, owned end to end.

I was the primary technical owner and sole repository CODEOWNER, from the first commit to production traffic in 28 weeks across 429 merged pull requests.

The intake half is multi-tenant ingestion of PDF, XLSX, and Google Sheets through Gemini extraction, confidence scoring, human review, and system-of-record delivery. Fourteen of fourteen validated production pushes landed with zero orphaned records, which is the number I care about most: the pipeline either delivers a complete job or it delivers nothing.

The semantic layer is mine end to end. Chunking and embedding workers cover 26,000 historical documents at 99.98% coverage and feed pgvector retrieval, so the agents answer from grounded source passages rather than from the extraction database alone.

On top of that sits a governed MCP server and agent architecture: 40 tools over 21 BI views, with tenant and principal isolation, audited queries, answers grounded in a cited source, and conversational surfaces deployed on Gemini Enterprise.

Extraction quality was treated as an engineering problem rather than a prompt problem. A golden-data evaluation harness wired into CI as a release gate moved quantity recall from 37% to 99%, and reached 96.3% precision at 91% coverage on a taxonomy mapping field that had resisted every heuristic before it.

The unglamorous half was making it a good place to work. CI runtime came down 52%, and the local unit test loop went from 10.5 minutes to 15.6 seconds.

Over a thousand vendor contracts ingested and indexed into a vector searchable knowledge base, cutting contract search time by 90%.

I built the MCP server that exposed semantic search and natural language querying over the corpus, along with automated portfolio reporting across spend, obligations, and risk exposure.

A $1M+ engagement where I led the team and owned the architecture through launch.

The piece that changed the client's economics was a schema-driven dynamic form engine: non-technical staff design their own intake workflows, without an engineer in the loop. It removed hundreds of hours of manual legal review a month.