AI-native engineer in Seattle. I build MCP servers, agent tooling, and test harnesses for systems that are hard to test.
Everything below is public and checkable: test counts, CI runs, registry entries, and commits in other people's repositories.
Stateless stdio MCP server, four read-only tools for observing Vercel deployments.
Batch-invariance verifier for llama.cpp continuous batching. Per-cell diffing, a five-verdict output contract, and signed GREEN or RED certificates. Mock-sourced passes are non-promotable by construction, so the gate cannot rubber-stamp itself. Ships a real Phi-4 RED divergence certificate so you can watch the gate bite.
Next.js 16 App Router app for tracking npm and MCP registry packages: download sparklines, server-paginated version history, a watchlist, markdown export, and explicit failure states including a rate-limit banner with reset countdown and stale-while-error handling.
Reference MCP server for cookie-session systems that expose no public API, with a Hono mock target and fixtures so the suite runs offline.
Escape Python reserved words in generated SDK identifiers. Covers every identifier site in the Python generator plus a bridge fix, preserving BAML wire identity while changing host-language spelling.
Resolve npx and uv ENOENT on Windows via shell spawn, with two platform-mocked tests.
A Grace Blackwell aarch64 launch-blocking shim, shipped default off so x86_64 does not pay the sync tax. I am one of two contributors to that project.
What BoundaryML/baml#4070 actually had to fix, and why escaping at one call site is not enough.
Cookie-session targets, an unstable HTML boundary, and where to put the stability layer.