Document data extraction that feeds theledger directly
A pipeline that reads incoming documents and pulls the fields the ledger needs, so data entry stops being a manual chore.
- Client
- Meridian
- Type
- AI
- Year
- 2024
- Delivered
- AI engineering · Extraction pipeline · Validation · Integration
- Method
- Vision model + validation
- Fallback
- Human review queue
- Output
- Direct to ledger
- Type
- Extraction pipeline
01 /
The challenge
Every document type had its own layout, and the same field could appear in different places. A template-based extractor broke on the first variation, and a wrong field keyed into the ledger is worse than no field at all.
02 /
What we did
- 01
Used a vision model to locate and read fields regardless of layout, so the extractor handles the variations a template cannot.
- 02
Built a validation layer that checks extracted values against expected formats and ranges, flagging anything suspicious for a human rather than accepting it.
- 03
Designed the pipeline to export straight into the ledger’s import format, so the last mile is a button, not a copy-paste.
- 04
Left a review queue where staff can correct flagged items, and fed those corrections back so the extractor improves over time.
03 /
Outcome
Document processing that used to take hours of keying now runs through a pipeline. Fields the system is confident about go straight in; the rest go to a short review queue. The ledger gets its data faster, and the team stops retyping.
Built with
- Python
- FastAPI
- OpenAI
- PostgreSQL