Turning fifteen thousand monthly PDFs into structured data

The challenge

Orbit received customs and freight paperwork from hundreds of partners, every one in a different layout, and four people retyped it into the operations system full-time. Template-based OCR had been tried twice and broke whenever a partner changed their form.

What we did

1. 01

   Used a vision-capable model to read documents by meaning rather than position, which removed the whole class of template-breakage failures.

2. 02

   Constrained every output to a strict schema, so downstream systems receive validated types or an explicit failure — never malformed guesses.

3. 03

   Attached a per-field confidence score and routed anything below threshold to a review queue, rather than treating all extractions as equal.

4. 04

   Built the review interface so a correction takes seconds and feeds directly back into the evaluation set.

The outcome

Field-level accuracy sits at 97.4%, with roughly one document in eight touching a human. Manual entry hours fell 82%, and the team that did the retyping now runs exception handling and partner onboarding.