Freight Bill of Lading Processing

A freight operator was processing 50+ bills of lading daily — averaging 4 pages each, some running to 80. Every document came from a different carrier, in a different format, many scanned at varying quality. They were doing manual spot-checks and missing discrepancies.

The actual problem

The core issue wasn’t OCR. It was that standard LLM pipelines treated each document as a single prompt — and on long BoLs, the model would simply forget line items from earlier pages by the time it reached the cargo summary. Weight and quantity mismatches were slipping through undetected.

The documents were long enough that context limits weren’t the problem. The problem was attention degradation across pages: the model reliably extracted fields from the current page, but lost fidelity when reconciling against totals declared elsewhere in the document.

What we built

A page-by-page extraction pipeline with rolling context: each page is processed individually, with a structured summary of prior pages passed forward. This keeps the model grounded across the full document regardless of length.

Extracted cargo data is then reconciled against declared totals — flagging mismatches in weight, quantity, and line item counts for human review. Not auto-corrected. Flagged. The operator stays in control of what gets resolved and how.

The pipeline handles variable scan quality and mixed carrier formats without retraining. Edge cases are caught at the reconciliation stage, not buried in a confidence score.

What changed

Discrepancies that were previously invisible to spot-checking became systematically catchable — across every document, every day. The team went from reactive error discovery to proactive mismatch review.

Volume processed: 50+ BoLs daily. Document length: 4–80 pages. Carrier formats: mixed, unstandardised. Human review load: focused on flagged exceptions only.