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

KineticLab turns an ordinary phone video of a swinging pendulum into auditable physics measurements. Meta's SAM 3.1 segments the bob and a reference card frame by frame (bob 591/591 valid, card 392/392 valid — pinned offline replay of a live capture); the deterministic, stdlib-only Python pipeline converts those masks into calibrated geometry (human-typed 85x54 mm card, reference type 'unknown' because a cut card is no standard, scale 0.40633 mm/px over 272 clean frames) and a pendulum fit reporting omega approx 5.38 1/s, period approx 1.168 s, and the identifiable g/L_eff = 28.93746036102984 1/s^2 with a seeded block-bootstrap interval [28.899, 28.983] 1/s^2 (n_boot 800, block 36 frames). The 591-frame, 19.73 s, 1024x576 clip keeps real source PTS timing; every export is hash-pinned; unknowns render as null — gravity stays null because no independent length was provided. Full suite: 877 tests, OK (skipped=15, platform-gated pins) (2026-11-03). Auditable, reproducible, honest: gravity is never claimed without an independently measured length.