# OnboardIQ

- **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:** [Pratham Rampurmath](https://cerebralvalley.ai/u/CipherPRG)
- **GitHub:** https://github.com/CipherPRG/onboardiq
- **Demo video:** https://youtu.be/qBE-j0qqIVo
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1087

Most AI guesses. OnboardIQ checks its own work on the real device.

THE PROBLEM
Every router, switch and server can report its health, but each vendor hides CPU, memory and temperature behind different IDs in thousand-line manuals (MIB files). Network engineers, MSPs and NOC teams read these by hand for every new device model. One wrong pick gives a dashboard that says "healthy" while the device is overloaded.

WHAT ONBOARDIQ DOES
1. Muse Spark 1.3 reads the device's manual and picks the object for each metric, predicting how it should behave.
2. Plain code proves every pick on the real device: it exists, has the right type and range, moves the predicted way under a real CPU or memory load, and agrees with an independent object.
3. A failed pick goes back to the model with the measured evidence, and it tries again. In the demo it rejects its own CPU pick (idle time, which falls under load) and replaces it with one the device proves.
4. Only proven metrics are monitored: a 0-100 health score, plain-English explanations from Muse Spark Open, and email alerts.
5. One click exports them to Prometheus and Zabbix, each checked against the real tool.

RESULTS
- Right object picked: 1% without the manual, 34% keyword matching, 73% with Muse Spark reading it (94 recorded real devices)
- 88% of picks that pass the device checks are correct
- A live onboarding took about 4 minutes and 5-27 cents in model calls

BONUS CATEGORIES
- Best Agentic: an autonomous read-pick-test-correct loop, where the device, not the model, decides what is true.
- Best Build on Meta Open Weight Models: Muse Spark Open, served on Baseten, writes the plain-English health explanations; code checks every number against the readings before it is shown, and a labelled template is used if the model is unavailable.

TRY IT
Hosted demo (no account, no key), recorded replays on GitHub Pages, or docker compose up from the repo. All runtime models are Meta models.

## More from Meta Global AI Developer Hackathon

- [BenchMate](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1084)
- [Velohud](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1085)
- [Vouch](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1086)
- [jobseekers](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1088)
- [CRYPTOR.AGENCY](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1089)
- [First Pass](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1090)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1087. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt