Weld Count

Weld Count automates the tedious process of weld takeoff — counting and classifying welding symbols across engineering drawing sets. A 100-page blueprint that takes an estimator two or more hours of manual review is processed in under five minutes. The platform serves welding engineers, fabrication shops, and estimators who need accurate weld counts for cost estimation and fabrication planning.

Technical Approach

The ML pipeline combines Detectron2 object detection with Tesseract OCR, both running on serverless NVIDIA T4 GPUs via Modal.com. PDFs are converted to 300 DPI images, OCR extracts text annotations using preprocessing tuned for engineering drawings (grayscale, denoise, Otsu binarization with sparse text mode), and the detection model identifies welding symbol locations. The detection threshold is set deliberately low to maximize recall — a missed weld in fabrication estimation is far more costly than a false positive that a user can dismiss.

The interactive document viewer, built with PDF.js and Konva.js, uses an immutable layer architecture: AI-generated detections and OCR data are never modified. User corrections are stored in separate edit files and merged at render time, preserving the original output for reprocessing and creating a clean audit trail. All coordinates are normalized to a 0-1 range for resolution-independent overlay across the wide variety of engineering drawing sizes (ANSI A through E, ISO A0-A4).

Key Decisions

Starting with single-class detection ("is this a weld symbol?") before attempting multi-class classification ("what type of weld?") — delivering immediate value while building the training dataset for the harder problem. Presigned S3 URLs for direct upload so large engineering PDFs never touch the application server. The immutable-original-plus-mutable-edits pattern enables model version comparison against the same documents as detection accuracy improves.