Hawkaii

Shoku reads a cooking reel into a recipe page: caption, comment, subtitles, then speech recognition, then sixteen frames of vision when the reel is silent — all structured by Muse into a schema-constrained recipe, cached once per reel. Then it acts on the result. A scoped assistant answers at the stove, proposes edits as a diff you explicitly apply, keeps version one, and turns "I have 2 kg rice and paneer" into fridge rows without inventing the missing amount. Suggestions are counted before they are chosen: the matcher does arithmetic, the model only ranks. In cook mode it listens through a server-side relay to Meta's streaming ASR, handles next/back/repeat/timers in three languages on-device, and speaks the rest. Model proposes, app disposes — with a meter on every answer.