Md. Tanjil Sarwar Lazim
Built at Meta Global AI Developer Hackathon · Oct 26, 2026 · Remote

OneShot checks whether a hands-on procedure was actually followed. Two people can pack the same box and end with the same final picture, but only one may have done the steps in the required order. A final photo can't show that. You record one correct demonstration. OneShot learns the steps, the final placements and which step orders matter: required (said out loud or physically necessary), as shown (a different order gets a note), or free (someone said so, or a second correct reference did it differently). A person confirms or edits the procedure. OneShot then checks later recordings for skipped steps, wrong placements and order mistakes, and points to the moment and the fix. Changing a rule re-checks the existing evidence without another model call. It's built for people who train staff or review packing and kitting work. When evidence is missing, such as a hand covering the box, it asks for review instead of guessing. These results are development evidence, not a held-out benchmark. On nine real phone recordings, six got the expected result, three went to review, and none of six incorrect attempts was accepted. On three generated order-only mistakes, a final-picture check accepted all three and OneShot accepted none. No second person has tested it yet. Live app: https://tanjilsarwar30--oneshot-web-web.us-east.modal.run