# Beyonder

- **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:** [Chandresh Gohel](https://cerebralvalley.ai/u/cg_)
- **GitHub:** https://github.com/Chandresh-Gohel/NASA_CMAPSS_Meta_Global_AIHT
- **Demo video:** https://www.youtube.com/watch?v=3SeoF6LADgI
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/750

Engine Health Agent helps maintenance engineers decide what to do with a jet engine, using NASA's C-MAPSS turbofan benchmark (simulated engines). Today's default inspecting every fixed number of cycles wastes hangar slots and still lets failures slip through: on our held-out replay it missed 42 of 106 failures while wasting 359 inspections. A standard remaining-life model doesn't solve this either: it emits a single number, silent about when that number is based on a faulty sensor or an engine outside its experience.

For each engine we predict remaining useful life as a calibrated 90% range, not a point. A multitask deep kernel learning model (CNN features into a Gaussian process shared across the four C-MAPSS fleets) is wrapped in conformal calibration so the band means what it says. Deterministic code never a language model converts the band into INSPECT, NO_INSPECTION or ABSTAIN. It refuses to call and routes the engine to an engineer when the band straddles the planning horizon, when a sensor fails integrity checks, or when the engine's uncertainty pattern is unlike the training fleet. Muse Spark agents explain each outcome: the driving sensors, the next step, and the reason for any abstention, with a safety reviewer and a hangar planner on top.

Replayed to failure on 106 held-out engines, the system missed none (0 vs 42 fixed-interval; 35 vs 359 unnecessary inspections), and every injected sensor fault and every foreign engine was escalated rather than answered confidently. The Streamlit app triages all 707 official test engines, supports fault injection and uploading your own engine file, and runs with or without an API key. This is decision support on a simulated benchmark, not an airworthiness determination.

## More from Meta Global AI Developer Hackathon

- [Omnesight](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/747)
- [Me and AI alone](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/748)
- [SMAJ Ecosystem](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/749)
- [binaryEXE](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/751)
- [Neuraforage](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/752)
- [AIBOSSPRO](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/753)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/750. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt