I build software with language models inside it, which means I spend most of my time on the
part your role is actually about: deciding whether a model's output is good, and being able
to say why.

Three examples from my own work:

- Compass — a local-first assistant that answers only from your own notes. It never
calls the model when nothing was retrieved, drops any citation that cannot be opened, and
keeps a turn on-device if any source memory is private. Rust + TypeScript, 235 tests.
- AutoApply — résumé tailoring where the model may only reorder and rephrase bullets the
user actually wrote. Every bullet cites its source id, and any number that cannot be traced
back to its source fails the document — a fabrication gate, not a disclaimer. Python,
FastAPI, 88 tests.
- A vision browser agent — Playwright plus a VLM, with a Pydantic-typed 13-action space
and set-of-marks screenshot annotation so the model selects an element id instead of
guessing coordinates.