# ASCII

- **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:** [Bekzatkhan Zhumabekov](https://cerebralvalley.ai/u/Beking003)
- **GitHub:** https://github.com/bekzatkhanDev/gridpilot
- **Demo video:** https://youtu.be/i2Tc22ZsykM?si=hdXKDIBZaEFSNI2r
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3455

GridPilot DC is a research prototype of an autonomous AI agent for diagnosing data-center infrastructure incidents. In real data centers, an incident rarely produces one clean signal: a rack overheats, power draw rises, but one sensor reads normal. An engineer who acts on the first alert may shift load before fixing the actual cause, and make things worse. GridPilot DC demonstrates a different approach: an agent that investigates before acting.

The system runs in a deterministic Next.js simulator with three racks, cooling zones, and conflicting sensor telemetry. A Muse Spark agent selects from three read-only diagnostic tools (inspect_rack, inspect_cooling, check_sensor_consistency), interprets results, and proposes a next step. A separate CAFC controller (Contradiction-Aware Fuzzy Control) decides whether there is enough evidence to act at all: ACT, INVESTIGATE, ASK, or STOP. A default-deny Policy Gate validates every proposed action against the scenario's allowlist and power headroom. An independent Verifier checks the post-action state. The agent can never mutate simulator state directly.

Scenario 2 — conflicting sensors — is the primary demo. Without a controller (muse_only), the agent acts prematurely 3/3 times. With fixed thresholds, 1/1. With CAFC, 0/3 premature actions and 3/3 verified outcomes. Preliminary results: N=3, raw counts, no statistical significance claimed.

Scenario 5 documents a known limitation honestly. When a phantom sensor masks a real problem in another rack, CAFC acts prematurely because CafcInput is per-rack-blind (measured: evidence, contradiction, and value of information are bit-identical between phantom and real-fix proposals). The next step — per-rack contradiction in computeCafcInput — is named and documented. ANFIS is documented as future work pending that input fix; we deliberately did not train on an input that cannot distinguish racks.

What this project does NOT claim: real infrastructure control, statistical significance, or a learned controller. It is a simulator-only research prototype. The submission's strength is a working guarded agent loop, a measured advantage on contradictory telemetry, and a boundary that we measured and named rather than hid.

Repo: https://github.com/bekzatkhanDev/gridpilot. Run it offline in three commands: npm i, npm test, npm run demo.

## More from Meta Global AI Developer Hackathon

- [Autumn's Second Bloom](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3452)
- [AVO](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3453)
- [TraceForm](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3454)
- [Amos Kanyanta](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3456)
- [Nearhand](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3457)
- [Sentinel AI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3458)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3455. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt