Bolt Finder originally assumed a combined barcode + OCR path would be the fastest way to identify rolls in warehouse rack imagery.
Testing changed that assumption. In real captures, barcodes degraded too much at rack distance because of glare, curvature, label wear, print quality variation, compression artifacts, and limited barcode resolution.
OCR became the primary path because it could still recover useful roll-number candidates from the same real rack images.
The current architecture is:
OCR → Candidate Extraction → Match Classification → Overlay Display
The working prototype now uses full-image OCR, tiled high-resolution OCR scanning, preprocessing variants, OCR fragment merging, candidate scoring, and exact-match/review classification.
Main lesson: in practical AI systems, success often comes from simplifying the workflow around the signal that survives real capture conditions.