Fastembed does inference well and model loading badly: it fetches whatever `main` points at, verifies nothing, and only knows its own table. So I kept it for inference and put a spec-driven loader in front of it.
How I verify Linux builds from a MacBook Air — offloading the compile to a Linux box (so the laptop stays cool) and catching deploy breakage before it ships, all composed within Nix.
How I got herdr-style pane indicators from ~100 lines spread across wezterm, fish and a pi extension — and why extending the tools you already run beats adding another tool on top.
How I built intelligent search with real-time syntax highlighting and autocompletion using CodeMirror and F#. Spoiler: the backend does the heavy lifting.