~/home / journal / 2026-07-15

Hayscript: empty folder to working product in one session

Local transcription with speakers, confidence flags, and markdown export. Nothing leaves the Mac.

Built a meeting-transcript app from an empty folder in one orchestrated session. Python engine on mlx-whisper large-v3 with speaker diarization, a React review UI, and clean markdown export at the end.

Speed on my Mac: a 100-minute recording transcribes in 17 minutes and diarizes in 4. All local, nothing uploaded anywhere.

Best moment of the build: the confidence panel flagged Whisper mishearing “Petani” as “Perthani” at 0.49 confidence, and I fixed it straight from the UI in one click. The tool caught its own mistake and handed me the pen. That is the entire product in one interaction.

cd .. cd ~