# WrenchVision

- **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:** [Lance Sidener](https://cerebralvalley.ai/u/moneybagsmcgee)
- **GitHub:** https://github.com/lancesidener343-netizen/WrenchVision-
- **Demo video:** https://www.youtube.com/watch?v=5eCYnWsfq-Q
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/365

I rebuild machines for a living. The part that costs the most time isn't the wrench work — it's finding the right manual. WrenchVision is a mobile web app that builds the manual around the job in front of you.
Snap a photo of the machine and describe the job, by typing or by voice. The app identifies the machine from the photo, then generates a repair manual: a Problem section up top (the reported problem and the most likely cause), numbered fix steps below it as the solution, a troubleshooting section with the check that confirms or rules out each likely cause, and a parts list with real manufacturer part numbers whenever they are verified — never guessed.
Every step carries a label — verified, inferred, or illustrative — so the manual never guesses and calls it fact. For a Honda EU7000is, steps the factory manual covers are built from that manual and labeled verified. More generally: when the app identifies a machine, it looks for that machine's factory owner's manual, and when it verifies one, it grounds the manual in it — steps the book backs are labeled verified, and the manual links to the book. When no factory book can be verified, every step stays labeled inferred, never dressed up as verified. The diagram area shows your actual photo with numbered gold markers on the located parts, with tap-to-zoom and part-name labels; when a part can't be located it gets no marker, and the diagram falls back to a labeled schematic. A spoken walkthrough reads the steps aloud, hands-free at the bench. Type extra symptoms or confirmations and the manual updates on the same job — it remembers the job. Tap a step done as you finish it, and the manual tracks what is complete, with green Done marks kept separate from the verified/inferred labels. The photo stays with the job — reload the page and the picture with its markers is still there. Jobs taken with no signal are held on the phone and sent when the network is back.
Main demo: Honda EU7000is generator. Also proven on a Casio watch, a MixAll C100 Pro live sound card, and a Cat 45 pump.
Built and tested for under $5 in API calls, on a $150 budget.

## More from Meta Global AI Developer Hackathon

- [ARIQ Water](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/362)
- [Saad Mansour · SOLANG](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/363)
- [thrqsytr](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/364)
- [Oddly Clear Studio](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/366)
- [LSE Group Corporation](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/367)
- [JCZ](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/368)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/365. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt