8 june 2026

monday

Spent today bouncing between polishing my public presence and tightening up an evaluation pipeline. A lot of it was “small details that matter”: cache keys, parameter plumbing, and build tooling edge cases.

What I learned

Cache keys need to encode all evaluation-relevant inputs. If an STT model/version isn’t part of the key, cached results can silently mix runs and break both resumability (after a crash) and comparability (across model versions).
Hugging Face datasets can download the full artifact even when you only iterate a small subset. Sampling (e.g., max_samples=10) limits iteration, not necessarily download size, so you need to plan around disk/network costs separately from evaluation time.
Claude Code access patterns are split across install and web. I dug through docs/search results to understand what’s available via desktop install vs “on the web” flows and what requirements/gating exist.

What I built

Updated my GitHub profile README: refreshed links (personal site), project blurbs, and the tech stack section to make the landing page more scannable and current.

Experimenting with

Eval pipeline fixes (work-in-progress):
Planning a change to include stt_model (version/name) in the evaluation cache key to make cached artifacts safe for resuming and model comparisons.
Threading new request/schema knobs like max_samples and stt_model through the API so they’re configurable instead of hardcoded.
Lining up an ORM field + Alembic migration for persisting the model identifier used for a run.
Electron/Vite packaging: ran pnpm + electron-builder builds and chased down a failed packaging/build output, mostly focusing on getting to a clean, repeatable release build.