Built a full record → transcribe → summarize pipeline from scratch. The core of the work
was recorder.py: getting WASAPI loopback to behave like OBS required three iterations —
blocking reads failed during silence, callback mode with gap reconstruction caused jitter,
and the final approach renders continuous keep-alive silence to hold the audio engine active
while streaming raw audio to disk and applying drift correction at the end. On top of that,
wrote transcribe.py (faster-whisper, auto GPU detection, portable CUDA lib loading, VAD
filter), summarize.py (map-reduce over the Anthropic API), and cleaned up watch_zoom.py
to auto-record Zoom meetings via process detection. Tied everything together with run.py
as a one-command entry point, setup.ps1 for one-time environment setup, run.bat as a
double-click launcher, and a full README with per-tool argument tables. Also handled
portability (no hardcoded paths, env-var config, .env + .gitignore), renamed main.py
to transcribe.py, and fixed a handful of Windows-specific bugs along the way (Unicode
crashes, CUDA DLL path, name collisions).