Skip to Main Content

FollowAlong

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

FollowAlong — Demo video

How-to videos and manuals can't see your hands. You only find out a step went wrong when the LED doesn't light, the cable doesn't work or the part is on backwards. Half of YouTube users say it's very important for learning things they've never done before (Pew Research Center), and nothing in that video knows whether you got it right. FollowAlong is instructions that see your work. Give it any how-to video, a paper manual (PDF) or phone photos of one, and it becomes a live, step-by-step guide that watches through your phone or laptop camera: - It checks each step before the next one starts, and loops the few seconds of the tutorial you need to see again. - It speaks up the moment something's off, like a teacher next to you: "Oh, wait, that short end's in front. Go behind the long part, then out to the right." - When it isn't sure, it asks you instead of guessing. Before a risky step (mains power, a car battery), it stops until you confirm. - It's hands-free: ask out loud, interrupt it, in 7 languages. - For teams and creators, it shows which step people get stuck on, and why. Who it's for: makers and students with electronics and robotics kits, people doing home repairs and installs, and assembly and workforce training, where a step has to be right the first time. It's not a chatbot wrapper. Instructions become a step plan, a camera watches the work, SAM 3.1 measures what's there, Muse Spark judges what geometry can't, and a measured failure always overrides the model. We built CF-Bench to test it with no human in the loop: 303 labelled frames from 9 Creative Commons tutorials. The combined system is the most accurate (70% vs 64% for the vision model alone) and recognises far more correctly finished work (68% vs 48%). Best case, a breadboard circuit: 93% of verdicts right and no unfinished step let through. Every model answer is in the repo, so every number re-scores offline. Try it in the browser with no key or camera: https://unnatibshah.github.io/followalong/ . Built from scratch Oct 26 – Nov 4, 2026.

Team