Hey, I'm
Staff Software Engineer building full-stack products and AI-powered systems.
I spent a day tuning settings to make a 27B model answer faster, and none of them worked. Switching to a 3B-active mixture-of-experts model made it 5–7× faster, and a 707-question check found it wasn't any worse.
Slim, distroless, alpine, or a compiled single binary — benchmarked on a dedicated amd64 host under cgroup limits. The base image turned out not to matter at all, and it took me two wrong results to find that out.
How I built system theme detection for my portfolio — an inline script for flash prevention, React hooks for reactive switching, and Shiki dual-theme syntax highlighting.
An offline GPS speedometer for iPhone. No account, no network, no analytics — the app has no networking code in it at all.
An OCR API that returns word-level geometry, not just a wall of text — plus the dashboard, API keys, and metered plans needed to sell it. One Go binary, SQLite, no external services in the request path.
Privacy-first AI scribe — record any conversation, get structured notes (SOAP, sales, 1:1, interviews). Audio stays on-device; AI runs in your browser, on the free tier, or against your own API key.