Skip to Main Content

Stephane Nguene

Built at Meta Global AI Developer Hackathon · Oct 26, 2026 · Remote

Stephane Nguene — Demo video

Tutorials tell you what to do. FixLoop checks that you actually did it. When your bike chain falls off, you watch a video, try, and only find out on the road that it wasn’t seated right. FixLoop is a repair agent that watches the repair through your phone and refuses to move on until it has measured that each step succeeded. Say “c’est fait” too early and it says no, shows the chain in red, tells you what’s wrong, and re-checks after you fix it. It ends with a 3-second spin test and logs the repair in the bike’s history. It is not an image analyzer: it runs a stateful loop (perceive → measure → decide → act → recover). SAM 3.1 segments chain and chainring every second, a geometric rule on the masks decides, and the language model is never allowed to validate. That rule comes from a measurement: Muse Spark alone judged a fallen chain “passed” with 0.82 confidence. Proof, not claims: a benchmark on 8 real clips, labelled before any run, compares FixLoop with “SAM rules alone” and “Spark alone”. FixLoop never read a fallen chain as seated (0 of 280 images). We also publish where it still fails: a hand lifting the chain can fool it. Try it with no install: https://wondrous-tiramisu-404f0a.netlify.app/

Team