All case studies / Case study 08 / 12
Brevane Logistics
Document extraction with a vision-language model
Scanned, handwritten or skewed documents turned into reliable data
In short
Bills of lading, invoices, delivery notes annotated by hand: a vision-language model reads each document the way a person would (layout, tables, stamps, notes) and turns it into structured, checked, searchable data.
- Client
- Brevane Logistics
- Sector
- Transport and logistics
- Size
- 600 staff, 9,000 documents a month
- −80%manual data entry on transport documents
- 96%of fields extracted with no correction
- 9,000documents processed a month
The context
Brevane Logistics receives thousands of documents a month: scanned bills of lading, carrier invoices in every format, delivery notes photographed on the dock and annotated by hand. A whole team was re-typing them into the management system.
The challenge
- 01
Classic OCR reads letters but loses the structure: it can’t tell which value belongs to which column.
- 02
Multi-column tables, stamps and handwritten notes break the usual tools.
- 03
A wrong weight or quantity ends in a customer dispute or a customs penalty.
- 04
The documents hold commercial data that must stay under control.
Our approach
Every document goes through a vision-language model that understands the whole image: it finds the header, tables, stamps and annotations, and outputs JSON that follows a schema defined with the business. Control rules check consistency (totals, weights, container numbers). The data goes to MongoDB, and the useful passages are indexed in Qdrant so they can be queried in plain words.
The solution, step by step
- 01
Hard documents
Skewed scans, dock photos, complex tables, handwritten notes: classic OCR gets lost.
- 02
Read like a person would
The model understands the layout: header, table, stamp, annotation.
- 03
Turned into data
Each document becomes JSON validated against a business schema.
- 04
Checked automatically
Inconsistencies (totals, weights, references) are flagged before any entry.
- 05
Searchable
“Which containers were delivered with reservations this month?”
Results
- −80%manual data entry
- 96%of fields extracted with no human correction
- 3 mininstead of a day to find a disputed document
- 100%of anomalies flagged before import
What we deliver
- Ingestion pipeline (scans, photos, PDFs)
- Vision-language extraction and business JSON schemas
- Control rules and a human review queue
- MongoDB storage and Qdrant vector index
- FastAPI API and plain-language search
Technologies
- Vision-language model (GPT-4o / LLaVA)
- Python
- FastAPI
- MongoDB
- Qdrant
- Docker
Timeline
8 to 12 weeks
Team
2 AI engineers, 1 back-end developer
Illustrative cases: clients and figures are fictional, representative of the solutions we deliver.