# Axi0N

- **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:** [Jeevith G](https://cerebralvalley.ai/u/JeevithG090), [Raunak Khandelwal](https://cerebralvalley.ai/u/R3unakk), [Navadarshanth Shanmugasundaram](https://cerebralvalley.ai/u/nav6x)
- **GitHub:** https://www.github.com/jeevithg090/axi0n_meta
- **Demo video:** https://youtu.be/jlMG_8AYTKw
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2802

Rulecam lets you tell the cameras you already own what worries you, in plain words and in your own language. Then it proves the rule works, checks every alert with a second Meta model, and keeps evidence anyone can verify.

The problem: warehouses, supermarkets and venues have cameras, but nobody watches them. One retailer found that less than 1% of its recorded video is ever viewed. Meanwhile, forklifts cause about 85 deaths and 34,900 serious injuries a year in the US, and crowd-crush inquiries keep finding that the warning existed but never turned into action. Plain-language camera alerts already exist. What a small site can't get is proof that a rule really works, that an alert is real, and what happens to the footage.

How it works:

* Say or type a rule, e.g. "keep people 4 metres from a moving forklift", in Spanish. Muse Spark 1.3 plans it: it surveys the camera, probes whether SAM 3.1 can really see what the rule needs, calibrates metres on the floor, compiles a typed rule, and says honestly what this camera can't check.

* Rule CI attacks every rule with edge cases and broken copies before a person approves it (the forklift rule passed 16 of 16 edge cases and caught 6 of 6 broken copies).

* Live, SAM 3.1 tracks the scene. When a rule fires, a second model looks at the frames before anyone is bothered. Workers can dispute any alert, and Muse Wonder reviews the dispute.

* Every step goes into a hash chain with a daily Merkle root that an inspector can re-check in their own browser.

* Muse Code writes new rule types, which must pass mutation tests and a Z3 proof before the owner can approve them.

One engine, three worlds:

* Warehouse: the Spanish forklift rule. The warning is spoken with Meta's MMS-TTS and heard back with Muse Voice Transcribe to make sure it says what it should.

* Supermarket: the owner asks the cameras anything, by text or voice. Spark picks its own tools (track queries, SAM counts, frames, video) and shows its reasoning. Spills and empty shelves become staff tasks, closed only by a photo.

* Crowd-crush early warning (recorded runs): per-zone head counts within 0-4% of the published truth on Jülich experiment footage, and a Spark warden that names the danger phase, writes announcements in Hindi and English, and checks a side exit isn't already filling before sending people there.

Staff get alerts by role on their phone. No face recognition, ever: alerts go to roles, never names, and faces are blurred in evidence.

Meta AI inside: Muse Spark 1.3 and 1.2, Muse Wonder 1.0, SAM 3.1, Muse Voice Transcribe, Muse Image, Muse Code and MMS-TTS. The live system runs on the Meta API alone, with no GPUs of our own. Earlier runs (1-4 Nov) also used Muse Glimmer 30B and Muse Spark Open on our own GPUs. Every number in the repo links to the report that measured it.

## More from Meta Global AI Developer Hackathon

- [Sotto](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2799)
- [Upgrade Gate](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2800)
- [JustMe](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2801)
- [herrye991](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2803)
- [Proof & Demo (solo: Dinesh Kumar Saravanan)](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2804)
- [ypahnu(solo)](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2805)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2802. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt