draftCheck

DraftCheck is an AI checker for engineering drawings. It catches GD&T and dimensioning errors (undefined datums, form controls that reference datums, missing basic dimensions or general tolerances, holes too close to an edge) before a drawing reaches the shop floor, where those errors turn into rework and scrap. I'm a mechanical engineer who spent two years building CAD software at Siemens (Solid Edge), and drawing checking is still largely manual. How it works: upload a PDF or image. Muse Spark Open reads it into structured data (every dimension, tolerance, datum and feature control frame, with locations and confidence), while Muse Spark reads it independently as a cross-check; disagreements are flagged for review. The engineer confirms or corrects the values. Six deterministic Python rules check the drawing and explain every pass or fail from the real values. An agent then proposes fixes with engineering reasons; Python applies each edit to a copy, re-runs the rules, rolls back anything that doesn't help, and caps the loop at three rounds, streamed live. The output is a redlined PDF (revision clouds, numbered deltas, handwritten corrections, a CHECKED stamp) and a CSV correction list. An "About this drawing" mode explains the part, its views, and each dimension's purpose for students. Core principle: the model reads and reasons; deterministic Python decides. Evaluation: 17 generated drawings with known answers across 10 part types (plates, a gear, fasteners, a flange and a shaft with section and detail views, a 3-view block, a bolted assembly). Each Meta model caught 11 of 13 seeded issues with 1 false alarm, and 10 of 10 fixes were verified by the rules; every miss is documented. Live at https://draftcheck.onrender.com, results dashboard at /eval.html, 831 automated tests with CI.