Been experimenting with OCR and AI document workflows lately.

One thing that surprised me: the OCR model usually isn’t the biggest issue in production.

Real users upload:

• blurry photos

• dark invoices

• cropped screenshots

• tilted ID cards

Simple preprocessing improved our extraction accuracy more than switching providers repeatedly.

Curious if others building AI/document workflows are seeing the same thing.