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

Mechanical Reviewer: from a drawing to a verified fix. Design-for-manufacturing (DFM) mistakes on mechanical drawings (walls too thin to machine, sharp internal corners, holes too deep for their diameter, tolerances tighter than the process can hold) are caught late, by a machinist or after a field failure. Mechanical Reviewer catches and fixes them at the drawing stage. Upload a 2D drawing and the agent: 1. Extracts a structured part spec with Muse Spark vision (JSON schema, with evidence boxes). 2. Checks it with deterministic DFM rules, then has Muse Spark re-read a zoomed crop of each issue to drop misreads. 3. Highlights each issue on the drawing with SAM 3.1. 4. Explains how each flaw fails in the field, backed by analogous real NHTSA recalls (local index of 1,067 mechanical recalls). 5. Fixes the design: Muse Spark proposes edits, code applies and validates them, re-checks, and feeds failures back (up to 3 rounds). 6. Proves the fix: the corrected drawing is re-rendered and re-read from the image; it must come back clean. Every model output is checked by code before it is trusted. On our benchmark of 25 drawings with 30 planted issues and 77 hard negatives: 100% recall, 100% precision, 99.8% extraction field accuracy, 95% fix pass rate and 95% clean re-verification. Built entirely during the event.