Every pillar, one tag. The range, not a skills list.
Transcription, diarisation, and speech running entirely on Apple silicon, and why keeping voice local is a product decision before it is a technical one.
In NutriM8 you can mumble your whole day into your phone once. A background worker untangles it into sleep, weight, exercise, hydration and food, and resolves "a snack after lunch" to a real timestamp.
Nikis needed a clean dataset of kids' holiday activities across NSW. So I pointed an agent at the open web and made it behave like infrastructure: typed, checkpointed, and capped.
Diarization is three questions wearing one coat. Split them into specialists, never merge across a turn, and the failures stop hiding inside each other.
A 600M-parameter local model beats Whisper Large v3 on a Mac. The backend choice is per-model, decided by latency and memory and GPU contention, not ideology.
Feeding long silent audio to a transcriber wastes time and money. Split on speech first using webrtcvad, then send only the chunks that contain voice.
An on-device speech stack for Apple Silicon, in Swift.
A recording goes in, a speaker-labelled transcript comes out.
Search food by meaning, not by exact product name.
A model drafts maths questions against the component library, a verifier throws out the junk, and a clean one renders. Forever.
whisper_schedule