# Anthony Cantu

- **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:** [Anthony Cantu](https://cerebralvalley.ai/u/anthonycantu)
- **GitHub:** https://github.com/ant10293/LearnLoop-Hackathon.git
- **Demo video:** https://youtu.be/MBAlZAZBJjg
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4948

**LearnLoop — Teach a task once. Verify it again.**

LearnLoop is an AI-powered application that learns physical procedures from video demonstrations and evaluates whether those procedures were completed correctly in subsequent recordings.

Today, verifying hands-on tasks often requires manual supervision or expensive, task-specific computer vision systems. LearnLoop explores a more accessible alternative: demonstrate a procedure once, review what the system learned, and use it to check future attempts.

**How it works:**

1. **Teach:** Upload a video demonstrating a task. Meta's Muse Wonder identifies objects, actions, and completion conditions, while SAM 3.1 tracks objects and DINOv2 builds visual appearance references.
2. **Review:** LearnLoop generates an editable, step-by-step procedure. Users can correct object labels, adjust requirements, establish ordering constraints, and approve the final procedure.
3. **Verify:** Upload a separate recording. LearnLoop independently analyzes the attempt and uses deterministic logic to compare observed evidence against the approved requirements.
4. **Inspect:** Receive step-level PASS, FAIL, CANNOT VERIFY, or NOT APPLICABLE results, with timestamps, visual evidence, and synchronized demonstration-versus-attempt playback.

Unlike systems that require an identical sequence of movements, LearnLoop can recognize successful completion even when steps occur in a different order, provided the procedure doesn't require a specific sequence.

When evidence is ambiguous, LearnLoop reports uncertainty rather than guessing. Targeted AI reinspection and human-confirmed observations provide additional ways to resolve inconclusive results.

**Demonstrated results:** LearnLoop has been tested using real fixed-camera and head-mounted recordings, including multi-step LEGO assembly and hardware tasks. In one demonstration, it learned a five-step motorcycle assembly procedure and correctly identified a missing component in a subsequent attempt. Development testing included 20 analysis runs, 16 verifications, and 75 live Meta API requests.

A built-in Reliability Lab also enables repeated independent analyses to investigate consistency and model variability.

LearnLoop demonstrates a path toward accessible, reusable visual verification for hands-on training, assembly, quality assurance, and other physical workflows—without requiring custom model training for each procedure.

## More from Meta Global AI Developer Hackathon

- [Dekay](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4945)
- [Mojulo](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4946)
- [The One Shot](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4947)
- [Persona Explainer](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4949)
- [Rosales Recommendations](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4950)
- [Schokolade](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4951)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4948. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt