# Muze AI Labs

- **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:** [Christopher Chavez](https://cerebralvalley.ai/u/Muze_AI)
- **GitHub:** https://github.com/muze-ai-consulting/hackaton-meta
- **Demo video:** https://www.youtube.com/watch?v=eGBVL64Ej6s
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/38

Second Look is an agent that stops scams inside the chat, at the moment someone is about to act.

Try it now: https://secondlook.muze.cl/ (live, real model, no login).

People reported losing US$ 15.9 billion to fraud in the US in 2025 (FTC). Scams work by creating pressure: a parcel fee, a bank alert. Most checkers ask you to stop and paste the message into a website, which is what people don't do under pressure. Second Look is for that person, and for the family who usually hears about it too late.

What the agent does. On a working WhatsApp Cloud API bot and on the web, you send a text, screenshot or voice note, and it works through four steps:
1. Decides. "Likely a scam", "Be careful" or "No clear signs of a scam", with the reason and next step. It never says "safe".
2. Investigates. Muse Spark picks its own tools: it follows a link's redirects without reading the page and looks up who registered the domain, even for links in screenshots.
3. Recognizes. It matches the message against 20 families built from 1,902 scams confirmed by Reddit communities.
4. Acts. A second agent decides what would protect this person and writes it: an alert to someone they trust, a report to their bank, blocking the sender. You approve each step. If the person already paid, the plan changes: the alert goes out at once and the report goes to their bank's fraud team.
The actions are simulated and labeled; the written report is real. The native WhatsApp view on our home page is a labeled concept.

Why you can trust the number. Second Look saw only what people posted on Reddit; the community's replies were the answer key. We used only posts from after Muse Spark 1.3 came out, wrote the pass rule first and committed the SHA-256 of all predictions before scoring.
- Attempt 1 (1,137 cases) did not pass: 0.787 balanced accuracy against a required 0.80. It's published.
- Attempt 2 (945 new cases) passed: 0.801 (95% CI 0.759 to 0.839), catching 84% of scams with false alarms on 23.8% of genuine cases. A keyword filter scored 0.517.
- The communities changed between attempts, so we can't credit the pass to our new question alone.
Recompute the main results with python3 scripts/verify_results.py (no key, no network). Limits are in docs/VALIDATION.md; voice notes and scammers trying to fool the model are untested.

In production: Docker behind nginx, about US$ 0.02 per check, spending caps on every model call, daily quotas, signed webhooks, no messages stored on disk. It's one server with no load test.

Built during the hackathon, with no prior code; first commit October 27. We started on audience simulation for ads, which failed its own backtest; Second Look was built from November 2 on. Old code remains in the repo, outside this product.

Not solved yet: a native version would have to be opt-in and keep end-to-end encryption; we haven't built that. The bot runs on the Cloud API test number, so it only reaches verified recipients until Meta verifies our business.

## More from Meta Global AI Developer Hackathon

- [HB Innovators](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/35)
- [paper-bro](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/36)
- [Ali Baba](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/37)
- [SellerPilot AI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/39)
- [ColonyIQ](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/40)
- [VisiLoop](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/41)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/38. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt