Keywatch

Proof is an AI back office for the many small merchants worldwide who run their business inside chat apps. They take orders by text, voice note and photo, get paid through local mobile wallets or bank transfers, and then match every payment to the right customer by hand. I know this problem firsthand from Bangladesh, where it is especially common, but the same steps repeat wherever sellers use chat apps and local payment systems, such as UPI in India. Payment systems plug in as adapters, so the same pipeline can serve any of them. Proof reads the customer's message, extracts the payment claim (transaction ID, amount, last 4 digits of the paying account), and matches it against the merchant's own payment feed. Plain code handles the clear cases: exact ID, one-character typo, amount plus last 4 digits plus time window, one payment covering two orders, partial payments and reused IDs. Muse Wonder is called only when plain code cannot read the message, for example a transaction ID typed with spaces or an amount written as "3 hazar taka". Proof never confirms on amount alone: when unsure, it asks. Every decision shows its reason and the payment record behind it, and confirmed orders go into a ledger. Results on my 38-thread synthetic benchmark: plain code alone got 33/38 correct (86.8%); with Muse Wonder as fallback, 38/38 with 0 wrong auto-confirms. Muse was needed on 9 of 38 messages. The benchmark and the payment feed are simulated and written by me, so these numbers show the pipeline works, not real-world accuracy. Not built: real chat-app or wallet integration, merchant accounts, and voice-note input inside the app.