# Nappers

- **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:** [Chethan Gowda](https://cerebralvalley.ai/u/behemoth), [Mayank Bhadrasen](https://cerebralvalley.ai/u/Mayank1024)
- **GitHub:** https://github.com/Mayank-1024/longtail
- **Demo video:** https://youtu.be/rPDpJZ9Zk8o
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1801

Self-driving teams need the rare moments to train and test on: a pedestrian half hidden by a parked car, a cyclist cutting in. Those moments are buried in hours of multi-camera, multi-sensor logs. Vision models can search video, but they say "yes" far too easily: on our held-out queries, 96% of the claims a vision model made on its own were false.

LongTail finds those moments from one sentence, and checks each one before showing it. You type a scenario ("a pedestrian partly hidden by a car"):
1. Muse Wonder turns it into a checkable scenario spec (actors, relations, durations).
2. A Muse Wonder agent decides which scenes to search and when to stop, within a per-query budget.
3. SAM 3.1 tracks each named object through all six cameras from a text prompt.
4. A deterministic LiDAR verifier checks every claim: is a real object in that box, at that distance, at that moment, and do the relations hold? Each claim is marked PASS, FAIL or UNVERIFIABLE, and only findings that pass are shown.
5. Muse Spark Open writes an evidence card. Every sentence must cite a measurement, and a validator rejects any that don't.

The web app streams the agent's trace and shows each finding across six cameras, a LiDAR bird's-eye view and a per-frame predicate timeline, with human approve/reject and JSON/CSV export.

Measured against PandaSet's human labels (15 held-out queries, 102 scenes, 95% bootstrap intervals), the LiDAR check raised precision from 0.088 to 0.20 and caught 59% of the remaining false claims. We report its costs too: recall dropped from 0.28 to 0.20, it wrongly rejects 46% of true claims, and the precision gain is not clearly significant. Three deliberate breaks (verifier off, calibration shifted, LiDAR out of sync) each made the numbers worse, so the evaluation can tell when the system is broken. Total spend on paid calls: $68.

## More from Meta Global AI Developer Hackathon

- [Chien Binh Chay Than](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1798)
- [SomePath](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1799)
- [Conan](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1800)
- [ItzDaiGaJseSami](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1802)
- [Stall to Store](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1803)
- [PerceptaHeal](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1804)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1801. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt