I build infrastructure for AI and developer systems: observability, evaluation, workflow validation, and the task queues underneath. Mostly Go and Rust.
Recurring theme across the work: make systems explain themselves. Why the bill moved, which deploy broke staging, whether an eval actually passed, which job died and why.
| Project | What it does |
|---|---|
| llmtrace | Self-hosted Go proxy that logs per-call LLM cost and latency, detects per-key spend anomalies, and runs an 8-tool Gemini agent to name the exact deploy that caused a spike. Live on Cloud Run. |
| spendlint | Pre-merge LLM cost gate for GitLab — reads each MR diff, projects the dollar delta against real traffic, and posts a verdict before it merges. The preventive counterpart to llmtrace. Live on Cloud Run. |
| evalc | Compiles a natural-language eval criterion into a deterministic Python evaluator — no LLM-as-judge — self-tests it in a sandbox, and registers it as an Arize Phoenix evaluator. |
| ACE | Rust CLI that runs multi-step API workflows as a typed state-machine graph: 21 static graph checks before any network call, plus cross-environment trace diffing. Shipped via Homebrew, Docker, GitHub Actions. |
| rivet | Postgres-backed task queue for Go. SKIP LOCKED dequeue, visibility-timeout crash recovery, at-least-once delivery. No Redis, no broker. |
BrainConnect-ASD — adversarial GCN with a Gradient Reversal Layer for site deconfounding on 1,102 ABIDE subjects (fMRI). Mean AUC 0.7872 across 20 leave-one-site-out splits. Fine-tuned Qwen2.5-7B for ASD diagnostic report generation. Ran on AMD MI300X for the AMD Developer Hackathon 2026.
Go · Rust · Python · PostgreSQL · SQLite · PyTorch · Docker · Google Cloud Run · AWS