BillSaathi

BillSaathi is an AI agent that checks an Indian bill or payment request before you pay. It answers SAFE, CAUTION or SCAM with the evidence, in English, Telugu or Hindi. "Your electricity will be cut tonight. Pay now." Messages like this reach Indian families every day, including people new to digital payments who may not read English. India lost ₹22,845 crore to cyber fraud in 2024 (Ministry of Home Affairs, Lok Sabha); fake bills are one common route. You send a bill photo, a message or a voice note, on the web or on WhatsApp. BillSaathi shows what the message says next to what the biller's record says, quotes the values behind each reason, and explains in your language, which it can read aloud. The product is the verification, not the reading. Code runs fixed checks first: is the biller registered, is every link and UPI ID official or a lookalike, does the wording match known scams, does the amount match the biller's record? A tool-calling agent on Muse Spark 1.3 reviews that evidence and may investigate further or ask one question. Its verdict must then pass deterministic rules and a second, independent Muse Spark call that checks every claim against its evidence. No model can turn a risk into SAFE: anything uncertain ends as CAUTION. On SAFE, the user can simulate a payment after explicit confirmation, with an idempotency key, a receipt and an append-only audit log. On SCAM or CAUTION, they get a warning card to forward to family, and every SCAM verdict shows the national helpline 1930. Measured on 21 synthetic test items held out from all tuning: 21/21 correct, 13/13 risky requests not called SAFE, 8/8 genuine bills passed, median 10.9 s, about $0.024 per check. A fresh 20-case challenge set, frozen before running: 20/20 verdicts. 73 fail-closed tests try to force SAFE with registry errors, prompt injection in three languages and invented links. The sets are small and written by me, and the /eval page says so. Privacy Mode runs the whole pipeline on the open-weight Muse Spark Open (via Baseten), with no call to Meta's Model API. Limits: biller records are simulated, not the live BBPS network, so a real bill ends as CAUTION by design. All data is synthetic and no real money moves. Next step: real biller data through a partner. Live app (one-tap samples, no sign-up): https://zealous-coast-09279dd00.1.azurestaticapps.net Evaluation: https://zealous-coast-09279dd00.1.azurestaticapps.net/eval WhatsApp is shown in the video; it runs on a test number limited to allow-listed phones. Built solo during the hackathon (Oct 26 – Nov 4, 2026) in a public repo; the commit history shows it. Coding assistant: Claude Code.