UpClose

UpClose: get up close before you buy. Used cars are usually bought after a quick walk around and a gut feeling, and the problems that cost real money (rust under the doors, patched panels, worn tires) are exactly the ones buyers miss. UpClose turns a phone walkaround into an evidence-backed condition report. How it works Capture: The buyer films the car in the mobile web app, either as one video or as guided clips. A three.js 3D guide shows where to walk and where to crouch for each of eight zones. Find the car: Meta's SAM 3.1 masks the vehicle in every sampled frame, so the parking lot, reflections and neighbouring cars are ignored. Find damage: SAM 3.1 runs again with seven damage prompts (rust, dents, scratches, paint bubbling, cracked glass, tire wear, fluid leaks), only inside that mask. Verify: Muse Spark 1.3 judges each candidate from its exact frame plus a crop, never the whole video. Every verdict is schema-validated before it's saved. Report: A deterministic, config-hashed rubric scores each zone. Every finding links to its evidence frame and timestamp. Zones that weren't filmed, or were too blurry, are marked "not checked" and never counted as clean. Engineering (FastAPI, PostgreSQL, S3-compatible storage, React) A Postgres job queue with fenced leases, so jobs are safe to run twice and a crashed worker can't publish stale results A budget ledger for every Meta API call: reserve → dispatch → settle Free and paid evidence scoped on the server Verified deletion of user media and everything derived from it 950+ backend tests and 130+ web tests Built solo during the hackathon (Oct 27 – Nov 4, 2026). All code in the repository was written during the event. UpClose is a visual screening aid, not a mechanical inspection. Model confidence is uncalibrated, and we make no accuracy claims.