futurpreneur

A kirana shop is run by one or two people standing for twelve hours, switching between customers, suppliers and a notebook. Inventory software assumes a keyboard, English, a barcode on every item and time to fill forms. None of that is true at the counter. Vocal Kaka replaces the form with a sentence. The shopkeeper says "do kilo cheeni kharida, ek sau bees rupaye" and the app transcribes it, parses it into structured ledger actions with Meta Muse Spark 1.1, matches or creates the item, projects stock, and answers questions like "aaj ka profit kya hai" — masking cost and margin when the active role is staff. Three things make it work in a real shop: Unit intelligence. Bori (50 kg), peti, crate, dozen, paav (0.25), savaa, paune, Devanagari numerals, and a self-learning unit registry for words the model has not seen. An append-only ledger. Every mutation is an event; a Postgres trigger projects inventory. Corrections are VOID + replacement, never an UPDATE on history. Retried voice commands are idempotent through content hashes and advisory locks — a replay returns the stored result instead of double-applying. Offline-first. Recordings queue locally and flush on reconnect, so a dropped connection at the counter costs nothing. Muse Spark is not decoration: it sits inside the transaction path. Its output is schema-constrained to the exact intent list the ledger validator accepts, and if the envelope is malformed the request fails over to Meta's Llama 3.3 70B open weights, then to gpt-oss as a reliability floor. Every response records which provider ran, so the claim is queryable in voice_command_logs.