Most recently
I built the flagship developer product at an AI platform startup: a 102,000-line desktop IDE, from zero in four months, holding one of two codebase-wide pull request approval seats.
Before that, sole architect and engineer of a 400,000-line recruiting intelligence platform on Azure. Before software, five years in the U.S. Navy operating P-3C acoustic sensor systems, tracking submarines.
- 95%of commits
- 1,085 of 1,142 on the flagship IDE, plus 262 merged PRs
- 102Klines
- Desktop IDE on Electron, from zero in four months
- 444endpoints
- Enterprise platform architected and built solo
- pfSenseemulator
- Modeled the GCP Cloud Router locally to validate BGP peering before the circuit existed
Six demos. All live.
Every API call is real. Nothing is mocked.
- Chess AI MagnusStockfish compiled to WebAssembly. Plays at full strength with no server round-trip.
- Research AgentPlans sub-questions, scrapes live sources, and cites every claim it makes.
- Financial DashboardLive market data behind a fleet-shared cache and an atomic daily spend ceiling.
- Reasoning LabDecomposes a question, argues both sides, and scores its own confidence against the evidence.
- Vision StudioObject detection and segmentation across four providers with automatic failover.
- Text to 3DPrompt to textured mesh, rendered and orbitable in the browser.
How it's built
The demos on this site run on the same footing as the production work: spend ceilings that refuse rather than overspend, database-level access control, and failure paths that say they failed instead of inventing a result.
- Multi-agent orchestration
- LangGraph
- pgvector + hybrid search
- Azure AI Foundry
- Event-driven, circuit-broken
- RBAC with tenant isolation
- BGP + colocated bare metal
- OpenTelemetry tracing
Selected writing
All posts →The parts that were actually hard: what broke, why, and what the fix cost.
- 13 part seriesHow to Architect an Enterprise AI System (And Why the Engineer Still Matters)A 13-part deep dive into the engineering decisions behind a production AI email intake pipeline, where the AI raises the floor and the engineer decides what to build on top of it.
- EssayValidation Geometry Is Part of the ModelA LightGBM capacity-control baseline for the ICAIF 2026 minute-ceiling paper that rebuilds its labels from parquet, and why label source, stride, purge, and temporal split geometry matter as much as the classifier in time-series ML.
- EssayThe Closed‑Loop Consistency Trick: Keeping Scene 12 Faithful to Scene 1 Without Global MemoryLong-form visual consistency looks like a memory problem, so I assumed it needed a growing story-so-far context hauled through every generation step. Scenematic works the other way around: propagate what must persist, measure what actually persisted, then correct what didn't, locally, with a periodic re-anchor that keeps small errors from compounding across a 20-scene run.
- EssayCaching LLM Extractions Without Lying: Conformal Gates + a Reasoning Budget AllocatorMy email extraction pipeline kept burning money re-deriving fields it already had. The bug wasn't a missing cache, it was caching without a validity model. What replaced it: a confidence-gated cache that decides reuse vs partial rebuild using conformal prediction, plus a reasoning budget allocator that spends compute per span only when quality is genuinely below target.