Bill of lading extraction for a freight operator — 50+ documents a day, up to 80 pages each.
Standard LLM pipelines treated each BoL as a single prompt. On long documents, the model forgot line items from earlier pages before reaching the cargo summary — and weight discrepancies slipped through. We built a page-by-page extraction pipeline with rolling context, reconciling extracted cargo data against declared totals and flagging mismatches for human review.
Read the case study