# RepairMind

- **Event:** [RAISE Summit Hackathon](https://cerebralvalley.ai/e/raise-summit-hackathon)
- **When:** Jul 4 at 9:00 AM – Jul 5 at 7:00 PM (GMT+2)
- **Where:** Paris, France
- **Team:** [Alexis Chapellier](https://cerebralvalley.ai/u/Jnkz)
- **GitHub:** https://github.com/Junkz3/deepfield
- **Demo video:** https://www.youtube.com/watch?v=fGyVvlguQbo
- **Gallery:** https://cerebralvalley.ai/e/raise-summit-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/raise-summit-hackathon/hackathon/gallery/41

Deepfield turns any document corpus into a grounded enterprise agent you can actually see thinking.

The problem: service organizations sit on thousands of pages of repair documentation (service manuals, scanned military TMs, wiring fold-outs, training videos) that no keyword search can read. A schematic has no useful text layer, OCR destroys tables, and technicians end up guessing.

Deepfield ingests any mix of PDFs, scans, photos and videos (drag a file onto the universe and watch it classify and join its constellation live), auto-organizes them into a knowledge base rendered as a navigable 3D universe, and calibrates an agent workspace over it. Retrieval is visual: VultronRetrieverPrime reranks the actual page images against the query (no vector store, no OCR pipeline), so scanned schematics and fold-outs rank correctly because they are seen, not text-extracted. A single NVIDIA Nemotron 3 Nano Omni model does all the thinking: planning, routing each request to the right specialist in a calibrated agent team, sufficiency judging (the agent decides on its own when evidence is insufficient and retrieves again), multimodal diagnosis over schematics and technician photos, and in-place page translation. The conversation is a real dialogue: the technician can ask anything about the corpus, and when the agent requests a measurement, a reading that contradicts its hypothesis flips the diagnosis with new citations. Every claim cites the exact page, table row or video second, confidence is an explained deterministic rubric, and each diagnosis compiles into a work order a service desk can use. The 3D universe is not decoration: it is the retrieval made visible. Probes light up candidate documents while the reranker scores them, cited pages fan out in space, and a WebXR mode lets you stand inside your knowledge base in VR.

Nothing in the engine is domain-specific. The Studio calibrates a new workspace from any corpus plus one sentence of intent: Nemotron designs the agent team and its operations from the documents themselves. The demo ships a second, insurance claims workspace born exactly this way. The flagship RepairCenter workspace covers 30 real manuals across 26 device types, from a Whirlpool dishwasher to a 1,704-page scanned HMMWV military manual.

Quality is measured, not claimed: 105 unit tests cover the deterministic engine, and a versioned benchmark of 39 hand-verified questions runs against the live agent (36/39 on the final build). Voice runs NVIDIA end to end: hold V and Parakeet ASR sends your words straight to the agent, verdicts are spoken by Magpie TTS, and the agent speaks 27 languages. The UI is touch-first down to iPads and phones. Everything runs on Vultr end to end: all inference on Vultr Serverless Inference, the app itself on a Vultr Cloud Compute VM; the API key never reaches the browser. The live demo is open to everyone: signup (email verification, per-account daily inference quotas) is open at https://deepfield.repairmind.io.

## More from RAISE Summit Hackathon

- [CareProof](https://cerebralvalley.ai/e/raise-summit-hackathon/hackathon/gallery/38)
- [Solveo](https://cerebralvalley.ai/e/raise-summit-hackathon/hackathon/gallery/39)
- [Team Rann](https://cerebralvalley.ai/e/raise-summit-hackathon/hackathon/gallery/40)
- [Xero](https://cerebralvalley.ai/e/raise-summit-hackathon/hackathon/gallery/42)
- [Team1](https://cerebralvalley.ai/e/raise-summit-hackathon/hackathon/gallery/43)
- [SOINAR](https://cerebralvalley.ai/e/raise-summit-hackathon/hackathon/gallery/44)

---

Markdown version of https://cerebralvalley.ai/e/raise-summit-hackathon/hackathon/gallery/41. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt