selected work.

I ship most of what I build in the open. Source and architecture both, so the work speaks before I do. The full profile lives on GitHub.

Inbox Triage Agent

Autonomous support-triage agent and evaluation harness on Claude Managed Agents. An 849-line SSE consumer with a three-condition terminal gate and reconnect with history replay, a 31/31 ship gate across ten grader-evaluated tickets, and 230 hermetic tests replaying production traces at $0 spend. TypeScript, MCP, Zod, Vitest.

Airlock

Local-first RAG document Q&A engine. Fully offline, with no framework, no vector database and no cloud API. Page-scoped chunking maps every citation to exactly one page, and retrieval runs as cosine similarity over a NumPy embedding matrix with zero network egress. Python, FastAPI, Next.js, sentence-transformers.

QuoteLens

Cross-platform AI field-quoting engine. A seven-node LangGraph pipeline turns site photos and spoken narration into itemized quotes, priced only from a fixed catalog so the model structurally cannot fabricate a price. 34 RLS policies across 9 tables. React Native (Expo), TypeScript, FastAPI, faster-whisper, Supabase.

Loadlane

Database-enforced multi-role isolation and proof suite. Four-role marketplace isolation as 10 row-level-security policies under FORCE ROW LEVEL SECURITY with a deny-by-default actor context, verified by a 13-case matrix that asserts the denial mechanism rather than mere failure. PostgreSQL 16, plpgsql, TypeScript, Drizzle, Vitest, Docker.

StemFoundry

Asynchronous ML audio source-separation platform. Demucs on PyTorch MPS, with GPU separation benchmarked at 0.28 to 0.58x realtime against a 0.89x CPU baseline, a compare-and-swap job claim race-tested to exactly one winner, and 130 network-free tests. Python, FastAPI, Celery, Redis, Next.js, Stripe.

VoxDesk

Real-time voice-agent scheduling and observability platform. Synchronous mid-conversation tool contracts for a voice agent booking real calendar events, per-tool time budgets derived from measured latency, and a 120-test suite that runs at $0 platform spend. Next.js, TypeScript, ElevenLabs Agents, Cal.com API v2, Drizzle, PostgreSQL.