SarafuPay (EPRIZE DIGITAL)

SarafuPay is how Tanzanian groups collect money together for weddings, funerals and community causes. Contributors pay by mobile money (M-Pesa, Airtel Money, Mixx) or card, and the organiser runs everything from WhatsApp in Kiswahili or English, by text or voice note. Live at sarafupay.app. During the hackathon we made the money trustworthy and verifiable with Meta's models: 1. Verifiable Trust Copilot. SAM 3.1 crops the amount and reference out of a receipt photo, Muse Spark reads each crop back and checks it against the ledger, and every decision becomes an Ed25519-signed, hash-chained evidence card that contributors can verify on a public proof page. 2. Swahili voice on Meta's open weights. Muse Spark Open listens to Swahili voice notes with about 40% fewer word errors than our previous transcription (measured on 12 real voice notes from 4 speakers), and replies with a spoken voice note from Meta's open Seamless M4T v2, which we host ourselves. 3. Kiswahili organiser copilot. Ask "nani bado hajalipa?" and Muse Spark investigates each payer across the ledger, webhooks and live provider status, then explains what happened. A verification layer outside the agent checks every amount, date and claim against the facts, rewrites failing replies with feedback, and flags anything it still can't verify. Reminders are drafted but only sent after the organiser says yes. 4. Personal thank-you videos. Muse Image draws a card in the collection's mood, Seamless M4T v2 speaks each contributor's name and amount, and every payer gets their own WhatsApp video. 5. **Muse Glimmer message router (fine-tuned open weights).** We LoRA-fine-tuned Muse Glimmer 30B on Fireworks with 687 Kiswahili, English and code-switched SarafuPay messages. On 97 messages from wordings it never saw, it went from 0% (base) to 90% exact answers, with 96% correct routing, in 0.7 s. Simple messages then get Muse Spark with only the tools they need: 56% fewer input tokens per call and 19% faster. Money messages always take the full path. Trained and tested; not yet switched on in production. What existed before Oct 27 and what is new is listed in HACKATHON.md in the repo.