# netmedic

- **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:** [Matteo Bianchi](https://cerebralvalley.ai/u/vmbr0)
- **GitHub:** https://github.com/vmbr068/netmedic
- **Demo video:** https://www.youtube.com/watch?v=_QdP5m4D6-Q
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/514

netmedic is an AI on-call network engineer. Network outages are still debugged by a person reading router output at 3 AM; the fix is often small (a filter typo, a wrong MTU, a pulled cable) but finding it takes an hour, and touching production without a safety net makes things worse. netmedic does the whole loop on its own, in about a minute:

1. Detect – it probes every router every 5 s (BGP, OSPF, best paths, reachability, DNS) against a baseline learned on the healthy network, so it also catches silent failures.
2. Investigate – Muse Spark runs a tool-calling loop over 14 validated, read-only router tools, guided by runbooks, and submits a structured diagnosis: root cause, evidence, a multi-step plan (one fix per root cause) and what must be escalated.
3. Fix safely – each step is a typed action with typed checks: config check, a timed apply that the router reverts by itself, verification, then commit or automatic rollback. Low-risk steps run alone; the rest wait for a human approval in the web UI.
4. See the physical world – when a link has no carrier and no config explains it, the agent itself asks the on-site technician for a photo. On the phone, SAM 3.1 segments every port live; Muse Spark vision reads the rack with the live symptoms and the cabling plan as context, and SAM cross-checks every "empty port" claim. The findings go back to the agent, which escalates the exact port to re-plug ("re-seat cable SW-MUL-TIU-LAB port 13" in the demo, on a real switch).
5. Report – every incident ends with a root-cause report and its full timeline.

Meta models: Muse Spark 1.3 (agent, tool calling, structured output, vision), SAM 3.1 (live port segmentation and verification of Spark's claims), Muse Wonder (last tier of an escalation ladder for incidents Spark cannot resolve).

Measured, not hoped: an eval harness replays 8 chaos scenarios on a lab of 8 real BIRD routers (NixOS VMs: OSPF, iBGP route reflectors, dual-homed upstreams, a branch). 15/16 runs resolved with no human, detection in 10–15 s, typically 1 minute to resolve, $0.002–0.014 per incident. The physical "pulled cable" scenario: photo requested and the right port escalated in under a minute, every time.

The agent loop is generic (probes → read-only tools → verified, reversible actions → a human for what it cannot fix): the network is its first domain; the same loop extends to other network OSes, Kubernetes and DevOps.

Everything in the repository and the video was built during the hackathon (first commit Oct 27).

## More from Meta Global AI Developer Hackathon

- [Shawn Ramey](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/511)
- [Camtoshop](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/512)
- [Smarties](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/513)
- [AI Playtest Studio](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/515)
- [Iris](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/516)
- [SiliconMind](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/517)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/514. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt