Automating freight-document processing for a D2C logistics platform
An AI-assisted document pipeline that cut manual review of freight paperwork from hours to minutes, with a human-in-the-loop safety net kept in place by design.
- Manual review time
- −81%
- Low-confidence auto-flag rate
- 100% routed to review
- Processing throughput
- 6× documents/hour
- Timeline
- 4-month engagement
The problem
Operations staff manually reviewed scanned bills of lading and customs paperwork to extract shipment data, a bottleneck that scaled linearly with order volume and routinely delayed time-sensitive shipments.
Our approach
We built an extraction pipeline combining OCR with a evaluated LLM extraction step, routed low-confidence results to a human review queue instead of auto-approving them, and instrumented the whole pipeline so accuracy could be measured, not assumed.
The outcome
The team redeployed two full-time reviewers to exception handling instead of routine data entry, while keeping a human decision on every shipment the model wasn’t confident about.
Stack
- Python
- FastAPI
- LLM evaluation pipeline
- AWS
- PostgreSQL