Document Intelligence Pipeline

An industrial group received thousands of documents a month — purchase orders, material certificates, delivery notes — as PDFs, scans, and email bodies. Every one was read by a person and typed into the ERP by hand.

Previous OCR attempts had failed on the long tail: rotated scans, supplier-specific layouts, handwritten corrections. Ninety percent accuracy meant someone still had to check everything, which meant nothing was actually saved.

We inverted the accuracy problem. Instead of trusting extraction, the pipeline scores its own confidence field by field; anything certain posts straight to the ERP, anything doubtful lands in a review queue showing the human only the fields in question.

Extraction combines layout-aware models with supplier-specific validation rules — a certificate that references an unknown heat number gets flagged no matter how cleanly it was read.

The system runs as scheduled pipelines the client’s own team can read, monitor, and extend. Handover included the runbooks.

Straight-through processing handles the bulk of inbound volume; the review queue needs one person for an hour a day instead of four people full-time. The team that used to retype documents now runs the system.