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

Wheelchair users often learn a route is impassable only when they reach it: the sidewalk ends, or the entrance has steps with no ramp and no handrail. AccessWalk turns a phone walkthrough video of a route into a know-before-you-go report. Anyone can film it: a friend, a caregiver, or the venue itself. SAM 3.1 tracks 14 kinds of objects through the video (steps, handrails, parked cars, bins, traffic barrels, raised slabs). A Muse Spark 1.3 agent then reviews the tracks with four tools and must cite a tracked object and frame for every confirmed barrier; a validator drops anything that doesn't check out. A second pass watches the whole video and adds "possible" findings, always marked verify on site. The report gives a route verdict, each finding with its evidence frame and box, a synced video timeline and a PDF. On 22 routes I filmed and labelled by hand (28 barriers), AccessWalk found 22 with one false alarm, versus 18 for Muse Spark 1.3 on the raw video alone. Faces are blurred and GPS stripped before any frame reaches a model. It's a screening tool, not an ADA certification.