Turns invoices, contracts and statements into structured JSON with 99.2% field accuracy and audit-grade traceability.
Trade finance lives and dies on document accuracy. The client's ops team was rekeying 30,000 documents per month across 14 issuer formats, and a 4% error rate was creating downstream reconciliation work worth roughly 18 FTEs.
Our pipeline combines Azure Document Intelligence for layout-aware OCR with a GPT-4 extraction pass guided by per-document-type JSON schemas. A deterministic validation layer enforces ISO codes, checksum rules and cross-field invariants before anything is committed. Every extracted value is rendered in the UI with a click-to-source highlight on the original page — auditors love it.
Field-level accuracy now sits at 99.2% on a held-out test set, and the team has redirected those 18 FTEs from rekeying to exception handling and supplier outreach. The pipeline is currently being extended to letters of credit and bills of lading.
A trade-finance team was rekeying 30k documents/month from 14 source formats with a 4% error rate that bled into downstream reporting.
We trained a hybrid OCR + LLM extraction pipeline, validated outputs with deterministic rules, and surfaced every value with a click-to-source highlight.
Designed for clarity. Built to scale.
Click-to-source field traceability
Deterministic validation layer
14 document types supported out-of-box
Snowflake & SAP export connectors
