Consai

Punchwalk turns a construction punch walk into a verified punch list: the list of defects that must be fixed before a job is closed out. I grew up in a construction family. Every job ends the same way: someone walks the site with blue tape and a notepad, then spends hours writing it up, and the final payment waits on that list. With Punchwalk, the inspector films the walk with Ray-Ban Meta glasses or a phone, says what's wrong, and points when it's subtle. Every item comes back with proof: the photo with the defect outlined, the exact second of video, and the inspector's own words, plus an AI summary, the fix, trade, room and severity. Verified items go out to each trade in Slack on their own; anything the photo doesn't back up waits for a human. On the re-walk, Punchwalk signs off what's fixed and catches false "fixed" claims. Three Meta models do the work: - Muse Voice Transcribe hears what was said and when. - SAM 3.1 follows the inspector's hand to the spot and outlines the defect. Punchwalk lines up each sentence with where the hand was pointing at that moment. - Muse Spark Open writes each item, locates each defect, and checks every outline and item against its photo before anyone sees it. On the re-walk, it compares the new footage with the original photo to sign off fixes and catch false "fixed" claims. All reasoning runs on open-weights Muse Spark Open via Baseten (meta-models/muse-spark-open-1026). On our labeled test walks: 12 of 12 defects found, 12 of 12 right trade and room, 2 of 2 false "fixed" claims caught. A 105-second walk becomes a punch list in under 5 minutes, not hours, for about $0.54. Next (Future): live capture straight from the glasses through Meta's Wearables Device Access Toolkit, so the punch list builds while the inspector is still walking instead of after an upload. Built from scratch during the hackathon. Try it at https://punchwalk.fly.dev