# Aga

- **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:** [Ahmad Garba Adamu](https://cerebralvalley.ai/u/rufatronics)
- **GitHub:** https://github.com/rufatronics/sentra-kariya-local-ai
- **Demo video:** https://youtu.be/HnrLB02GhDw?si=GCw18u65-aYTUWes
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3358

Sentra helps people in Nigeria check suspicious documents and messages, in Hausa, English, Yoruba and Igbo, with everything running on their own device.

The problem: fake documents and scam messages (fake alerts, "you won" SMS, requests for BVN or OTP) are common, and most people cannot easily check them. Existing tools need fast internet, rely on cloud APIs, and rarely support Hausa. Many users have limited data and older devices.

What it does:
- Document check: the user uploads a photo of a document. Forensic analysis runs in the browser (error level analysis, frequency analysis, noise consistency, text-line alignment, compression history) and produces a trust score and a plain report.
- Scam check: pasted messages and links are checked for common scam patterns.
- Plain-language explanations: Meta's Llama 3.2 1B writes a short explanation of the result.
- Local-language results: Meta's NLLB-200 translates results into Hausa, Yoruba and Igbo.

Design choices: the verdict is calculated by deterministic code from the forensic signals, and the small language model only writes the explanation, so a 1B model cannot change the verdict. If the device has no WebGPU or the user skips the model download, the app still works in a basic mode with forensic checks and rule-based results. Models download once from a setup screen, then work offline, and no document leaves the device.

## More from Meta Global AI Developer Hackathon

- [Human](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3355)
- [Hyper](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3356)
- [DockTwin](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3357)
- [MuseOrg](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3359)
- [Cutting Edge: A Ninja Fighting Game](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3360)
- [PollPal](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3361)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3358. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt