OCR Document Processing System

The problem

The brokerage processed thousands of freight invoices, bills of lading, and customs declarations every week, most arriving as scanned PDFs or photographed paperwork from drivers. A data entry team re-typed the key fields into their internal system by hand, and mistakes in things like weight or HS codes caused downstream billing and compliance issues.

What we built

• Built a document pipeline that classifies each incoming file by type before extraction, since invoices, bills of lading, and customs forms all have different layouts.
• Combined OCR with a layout-aware extraction model so the system understands which numbers are which — a total isn't confused with a tax line or a tracking number.
• Added a confidence-scoring step: high-confidence extractions post straight to the internal system, while low-confidence fields are flagged for a two-minute human check instead of a full manual re-entry.
• Fed corrected fields back into the extraction model on a regular cadence so accuracy kept improving as document volume grew.

The result

The system now handles over 10,000 documents a day at roughly 98% field-level accuracy, and the data entry team shifted from full manual re-keying to spot-checking the small fraction the system flags — cutting manual entry time by about 95%.