# Unmanually

- **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:** [Moni Mindich](https://cerebralvalley.ai/u/unmanually)
- **GitHub:** https://github.com/gws-builds/show-me-the-problem
- **Demo video:** https://youtu.be/WFk0nOupw6c
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1843

Show Me the Problem: a live call that sees, for plumbing, HVAC, electrical and appliance repair businesses.

Homeowners often can't describe what's broken, so the office guesses. That costs trade businesses trucks sent for problems a homeowner could fix in a minute (pure cost for companies on fixed-price service contracts), second trips for the wrong part, phone tag, and hazards nobody saw until the technician arrived. Here the homeowner points their phone at the problem during the call.

The agent: Muse Spark 1.3 runs an agent with 10+ tools (find a part, direct the camera, inspect the frame, read a rating plate, record a hypothesis, give water shut-off steps, quote, request a booking). SAM 3.1 outlines the part on the live camera, and code turns the mask into "tilt up, closer, hold still". Muse Voice Transcribe listens in real time, in English and Spanish. It also leads a guided tour of a boiler room or electrical panel, naming every part with a status.

What code decides, not the model: prices come only from the owner's list, and only when SAM saw the part; a spoken hazard stops quoting before any model call; a camera hazard needs Spark and SAM to agree; nothing is booked until the owner taps approve.

After approval: Muse Wonder writes the technician's order of checks, every step citing call evidence. Muse Glimmer 30B, open weights on a GPU the business controls, drafts the cited truck list and the confirmation text.

Measured: an eval of 61 cases on public photos: 60 passed, 10/10 spoken and 3/3 camera hazards escalated, 0 false escalations, 5/5 prompt-injection attempts changed nothing. Four recorded calls: $0.13–0.26 of model spend each, a median of 6.0 s to a useful answer, and the microwave quote was ready 1.6 minutes into the call. 263 tests; with no key the repo replays the four real calls.

Try it live (hosted copy for judges): https://gwsbuilds--snap-hosted-serve.us-east.modal.run/k/judge-pumd5mneq8egskkj94zc?to=go . Open it on a phone, or on a laptop with a camera, and point it at an appliance, a sink or a panel; the owner page approves the booking. Spend is capped in code, and the after-booking step can take a few minutes on a cold GPU.

What's new: all code was built Oct 27–Nov 4, 2026. Our existing product, Unmanually, an AI office manager for trade businesses, is not in this repo; this is designed to ship into it as a video-call step. The demo business is fictional.

## More from Meta Global AI Developer Hackathon

- [Gol](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1840)
- [Spark Gate](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1841)
- [WorldKit](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1842)
- [Trust Boundary](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1844)
- [Akshara](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1845)
- [NIMICODER](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1846)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1843. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt