# Cuelga

- **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:** [Sebastian Ferral](https://cerebralvalley.ai/u/SEBASS)
- **GitHub:** https://github.com/sebassfc/cuelga
- **Demo video:** https://youtu.be/vsCWR6gyoSM
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5298

Cuelga ("hang up" in Mexican Spanish) protects families from virtual kidnapping calls while the call is still happening.

The scam: someone calls grandma, a voice cries "¡Abuelita, ayúdame! ¡Me tienen!", a fake commander keeps her on the line ("don't hang up, don't tell anyone") and demands a deposit at OXXO. The relative is almost always safe, but the person on the phone can't check without hanging up. In 2025 Nuevo León, my state, recorded 991 extortion victims (SESNSP data via Grupo Animal).

How it works: the family sets up, with consent, who might be "kidnapped" and a second contact. When a suspicious call comes in, you put it on speaker next to Cuelga. Muse Spark Open, Meta's open-weights model, listens to each audio chunk, transcribes it, tells the caller from the person at home and tags the stage of the extortion script (emotional hook, isolation, fake authority, demand, payment). As soon as the script shows up, an agent silently sends the relative a push notification, "¿Estás bien?", while the call keeps going; the phone can stay locked. If they tap yes, the home screen turns green: "ANA ESTÁ A SALVO. CUELGA." If nobody answers, it asks the second contact, then tells you to hang up and call 089, Mexico's anonymous reporting line. It never advises paying or negotiating.

Safety by construction: a Mexican-Spanish lexicon always runs as a floor; the model can only raise the alarm if it quotes the call word for word; everything the victim sees comes from a closed message list; the agent policy is property-tested; spend caps are enforced in code.

Try it in one minute, no setup, with English subtitles: https://cuelga.onrender.com/demo

Measured so far: on the live demo the scripted call alerts at 2 s and confirms "safe" 6–18 s after the first line, depending on how fast the relative answers, for about a cent of model calls. The live ear transcribed a 23 s synthetic-voice call in 8 s chunks (4.9–8.9 s per chunk, $0.026). On 8 synthetic scripts: 4/4 scams caught, 0/4 false alarms. These are synthetic tests, not real victims. Known limits are in the README.

Everything was built during the event, starting Oct 29, 2026.

## More from Meta Global AI Developer Hackathon

- [TechZ](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5295)
- [UNISON](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5296)
- [New team](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5297)
- [Quien Tiene (Axon)](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5299)
- [Deskverse](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5300)
- [KushZombie](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5301)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5298. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt