I build WDO Desk — software that turns a termite inspector's spoken walkthrough into a structured, review-ready state inspection report. Live in production with a California WDO operator since July 2026.
- Captures a field inspection by voice and diagram, offline-capable, on a tablet
- Extracts findings in the regulation's own terms (CA WDO Form 43M-41) — the model is structurally constrained to the regulation's taxonomy, it can't invent a finding code
- A person reviews every finding before anything is filed — nothing files itself
- Every correction becomes a permanent test case the system has to pass from then on
- Multi-tenant platform (internal name: Anvil), schema-per-tenant isolation, Postgres + Drizzle, deployed on Railway and Cloudflare
- An eval-gated, self-improving dev loop: failing test case → code → full gate (lint, types, tests, evals) → PR behind automated + human review
- The moat is provenance and auditability, not speed — every finding traces back to its source capture and model version
10+ years building interfaces and leading engineering teams. Former Chapter Lead at Airbyte; led platform modernization and an 8-person team at HouseCanary.
Open to consulting engagements in applied AI and regulated-domain automation.