# Kavyansh Bagotra

- **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:** [Kavyansh Bagotra](https://cerebralvalley.ai/u/itskavyansh)
- **GitHub:** https://github.com/itskavyansh/Canary-Analyst
- **Demo video:** https://youtu.be/t0V-uzax8CY
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3027

Canary Analyst is an AI layer for ProjectCanary, a ground-subsidence early-warning concept for areas above abandoned underground coal workings. Rule-based alarms fire on harmless events such as rain tilt, blasts, machinery impacts and sensor faults, and crews learn to ignore them. Canary Analyst investigates each alert with tools and recommends KEEP or DOWNGRADE. A deterministic guard can veto any downgrade, so the AI proposes and fixed code decides. The system only recommends. Humans decide on evacuation.

On a held-out test set of 300 simulated 6-hour scenarios, false-alarm scenarios fell from 24 to 16 (8.0% to 5.3%), with 8 removed and 0 added (paired sign test p=0.0078), while hazard detection stayed at 86 of 86. The confidence intervals overlap, and most of the reduction from the raw rules comes from the frozen rules baseline; the agent adds those 8 removals on top, each checked by the guard. The guard's final action was KEEP on all 110 hazard candidates. On one of them, the analyst recommended downgrading a real hazard and the guard vetoed it. All 714 evidence IDs the agent cited were valid. Silent sensor faults remain unresolved: 12 of 43 still alarm.

An audit after the test run found a cache-key bug that served 16 test candidates from earlier dev-split entries. We re-ran them live on test data: all stayed KEEP, one label changed from unclear to ground_movement, and detection was unchanged. We corrected the published candidate-level numbers and listed this and other known issues in the README.

The demo also replays one 37-minute real recording from a 2-node bench prototype. It is a demonstration, not a benchmark, and no labels were given to any detector. All benchmark results use simulated data, and thresholds are demonstration values.

A Live Sandbox tab lets anyone design a synthetic scenario with sliders; it runs the same frozen detector and, if it alarms, calls Muse Spark 1.3 live with the same prompt, tools and guard. Each run is a single synthetic test, not a benchmark result.

Live demo: https://itskavyansh.github.io/Canary-Analyst

## More from Meta Global AI Developer Hackathon

- [FixScribe Team](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3024)
- [Walkout](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3025)
- [Debate AI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3026)
- [Ghostline](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3028)
- [Ernest Pascual](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3029)
- [Hudson Forge Technologies](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3030)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3027. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt