Forward Deployed AI Engineer
I build production LLM systems on the Anthropic API and the Claude Code Agent SDK — taking them from scoping through prototype, evaluation, hardening, and handoff. I spend most of my time on the unsexy parts: structured tool use, retrieval that doesn't hallucinate, evaluation harnesses that catch regressions before users do.
One authenticated HTTP/SSE control plane fronting 100 third-party services (37 HTTP + 63 stdio). Server-side credential injection with auto-refresh, role-based policy, per-call audit logging, semantic discovery via OpenAI embeddings + Supabase pgvector.
Bun · Hono · Supabase pgvector · Fly.io
Plain-English LLM interface that turns natural-language requests into confirmed, audited actions across 69 tools. Confirm-before-mutate safety gate, model-cascade auto-close, full audit logging.
Next.js 16 · TypeScript · Supabase · OpenAI + Anthropic · MCP · Fly.io
Browser getUserMedia → JWT-gated Hono endpoint (MIME allow-list, 25 MB guard, full HTTP error taxonomy) → OpenAI Whisper-1. Unit-tested both ends.
React 19 · Next.js 16 · Zod · Bun
Delivery infrastructure on Fly.io for 12 repositories with an 8-minute build SLA. Autoscaling dispatcher, drain-aware pools, two-tier cache, 30+ reusable workflow composites, fail-closed merge gate.
Fly.io · GitHub Actions · Bun composites