# Prudence

- **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:** [Hadrien Baradel](https://cerebralvalley.ai/u/Takeo)
- **GitHub:** https://github.com/hadrienbaradel/scam-shield
- **Demo video:** https://youtu.be/XdCqM1nhpfU
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3558

Americans over 60 reported $7.7 billion lost to online fraud in 2025 (FBI IC3). The scams that get past the spam filter arrive at 9 PM, when nobody is there to ask. Prudence is the person to ask.

It connects to Gmail and checks every new message against the senior's own life, learned from their genuine mail: the bank and companies that really write to them, recent orders and carriers, the people they write to, data breaches they were part of. Texts, voicemails and pictures can be checked too. Each check ends in one verdict (Scam, Be careful, Safe), one action, and the reasons in plain words, read aloud on request: "USPS never asks for a fee by text. Your order ships with Amazon Logistics, not USPS."

Meta's models do the judging. Rules gather the evidence (sender, links, her context), Muse Spark decides and explains in a fixed JSON schema, and a guardrail stops it from clearing a proven scam. Prudence never sends, never deletes: a scam goes to a Scam folder in her Gmail, and Undo brings it back. A trusted person (her grandson) can be asked in one tap: her own phone writes the text, and his answer appears on her screen. Personal numbers are masked before Meta sees anything, emails are never stored, and Muse Spark runs in Standard mode.

Measured: on 18,849 public test emails, Muse Spark lifts phishing caught from 29% (rules alone) to 81%, with 1 false Scam in 17,291 genuine emails. On Hadrien's own inbox, Muse Spark judged 430 emails and called none of them Scam by mistake. Both scams in the video reached Hadrien the week of the event: a fake Norton invoice hidden as a picture in a spreadsheet, and a "free cruise" voicemail.

Built also with Vlad Zloteanu as part of the team.

## More from Meta Global AI Developer Hackathon

- [akivaron](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3555)
- [gentcod](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3556)
- [Team Lumen](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3557)
- [itlogs.food](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3559)
- [Altair](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3560)
- [Simulicity IA](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3561)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3558. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt