Doc AI Museum: Room 1

Back to the post

This room shows why early document AI systems could work well on narrow document families, and why they broke so easily. The workflow is simple: register the incoming scan against a template, crop expected field regions, and run OCR on those regions.

The failure mode is not abstract. A template stack can still look solid after registration succeeds. What breaks next is usually the fixed crop plus OCR step, especially on small fields and multiline content.

The assets here come from a real offline generator in the same folder: synthetic employment-application forms, feature-based image matching, DocTR OCR, and a hybrid corruption pipeline that simulates print defects, paper defects, geometry drift, and capture artifacts.

What To Try

• Start on Clean. The full pipeline works: 4/4 fields extracted correctly.
• Click Rotation, Warp, or Phone photo. Registration still succeeds but OCR quality varies.
• Click Blur, JPEG, or Phone photo. Matching still works, but OCR starts degrading on small fields.
• Now click Revised form. Registration succeeds with more inlier matches than clean. But the form has one fewer employment-history row, so Section 3 shifted up. The template’s crop coordinates silently miss Records due (days). This is the failure mode that motivates field detectors.
• Click a field card to inspect the crop region.