# ActCheck

- **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:** [Miras Auganbayev](https://cerebralvalley.ai/u/miras_auganbayev), [Atilla Kenebay](https://cerebralvalley.ai/u/Atilla)
- **GitHub:** https://github.com/Atilla888/ActCheck
- **Demo video:** https://youtu.be/s0BQPzDzUNY
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2720

ActCheck checks a renovation bill against photos of the finished work, before a family pays.

Families pay contractors from a work statement: "Wall tiling, 5.6 m² × 9,500 KZT". Most cannot verify that number, and a tape measure does not settle the argument. Tile is laid on a regular grid and its size is printed on the box, so the room is its own ruler. No reference object, no depth sensor, just an ordinary photo.

How it uses Meta's models:
- SAM 3.1 works as a measuring instrument. It segments every tile on the full frame and nine overlapping crops, and also finds doors, windows, mirrors, bathtubs and the floor. Tile masks are fitted to a lattice, which gives a homography from the image to the wall plane in millimetres, and an area with its uncertainty.
- Muse Spark 1.3 reads photos and PDFs of statements in Russian and Kazakh into structured lines.
- Muse Image generated the synthetic walls in our demo.

Models read and segment; ordinary code decides. Arithmetic, geometry, comparison and every verdict are deterministic, so a result can be traced. A line is flagged only if billed > measured + 2σ + hidden area. A wall that is not fully measured gets "needs more evidence", never an accusation. Hidden work is not guessed: it becomes a question. Reports come in Russian, Kazakh and English, with polite questions for the contractor.

Results, all on synthetic data:
- SAM 3.1 on 16 renders with exact truth: 11 reported as measured, median error 0.89%, maximum 2.44%, all within 2σ; the other 5 flagged or refused. This run predates our colour check and was not repeated.
- Planted inflation with simulated masks: 0% false alarms, 58% of 20% inflations caught. With SAM on renders: 1 false alarm in 15 honest bills.
- 5 synthetic statements: every checked field matched, 2 of 2 planted errors found.

Limits: accuracy on real walls is not established, because no wall has been compared with a laser or tape yet. On real interior photos the wall often does not fit the frame and is reported as partial. Only wall tiling is measured.

The bills and walls in the video are synthetic, and the family in it is illustrative. The analysis in the video is a real run, sped up on screen.

There is no hosted version: from the cloud servers we tried, the Meta Model API returned an empty model list and 404, and we have not established why. Run it locally with your own key using Docker; the steps are in the README.

Built during the hackathon by Atilla Kenebay and Miras Auganbayev. AI coding assistants helped write the code; dependencies and the timing of the first commits are disclosed in the README.

## More from Meta Global AI Developer Hackathon

- [BioDigit](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2717)
- [Layer_Zero](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2718)
- [Barghest](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2719)
- [Violet Sky Security](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2721)
- [Quack](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2722)
- [Aurora](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2723)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2720. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt