# precious

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Precious Adepoju](https://cerebralvalley.ai/u/zev)
- **GitHub:** https://github.com/preciousadepoju/alertcheck.git
- **Demo video:** https://youtu.be/SkJ5m8hr21E
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2888

AlertCheck protects Nigerian sellers from "fake alerts", a scam where a buyer shows a doctored or fake "transfer successful" screen, takes the goods, and the money never arrives. Police and fact-checkers give the same advice: confirm the money actually reached your account. That's hard to do with a queue waiting and a buyer insisting, and a shop attendant often has no access to the owner's bank account at all.

A seller or attendant sends the buyer's receipt to a Telegram bot, with the price as the caption. AlertCheck replies RELEASE, DON'T RELEASE or CHECK MANUALLY, with a one-line reason and a box around the field that doesn't match, such as "Don't release yet: only ₦5,000 of ₦50,000 has arrived."

The key design choice: the model reads the receipt, and the ledger decides. AlertCheck doesn't claim to spot forgeries from an image. Every verdict rests on whether a matching credit exists in the seller's incoming payments, and deterministic, tested Python rules make the final decision. That lets it catch tricks a balance check can miss: a small real payment with a receipt claiming the full amount (the "double alert"), an old genuine receipt reused for a new sale, and scheduled or pending transfers. After a sale, /sold marks the payment as used so it can't be used again. Owners can invite attendants who can check receipts without seeing the shop's payments.

Three Meta models power it:
- Muse Spark 1.3 reads receipts into structured fields and runs a tool-calling agent that searches the ledger.
- SAM 3.1 crops photos of a phone down to the screen.
- Muse Voice Transcribe lets sellers ask by voice, such as "did the fifty thousand from Chidi come in?"

On our test set of 42 synthetic receipts covering each fraud pattern, run end to end on Meta's models, AlertCheck released 0 of 24 fakes, wrongly blocked 0 of 14 genuine payments, and averaged 11.5 seconds per check.

Limitations: the demo uses a separate ledger of synthetic payments, not a live bank connection. An SMS credit-alert feed is built but hasn't been tried on a real phone, and a direct bank link is the next step. Real receipts are messier than our test set, so real-world accuracy isn't measured yet.

The Telegram bot is private because it handles payment checks. The README shows how to run AlertCheck without an API key, using synthetic data.

## More from Meta Global AI Developer Hackathon

- [khaled ehealth](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2885)
- [Evarest Technologies](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2886)
- [Koblip](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2887)
- [Evangelina Hoepfer Tinkers Space](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2889)
- [Hot Seat](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2890)
- [Ghost Run](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2891)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2888. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt