The problem to explore
What I want to understand.
Reduce repetitive classification work while preserving the context and human judgement needed for reliable accounting decisions.
Planned approach
Where to start.
Extract document structure, generate line-by-line suggestions using context, history and confidence signals, and require human validation before any record is created.
Next steps
From idea to experiment.
- Define the document extraction model
- Design the suggestion contract
- Prototype line-level classification
- Add an explicit human validation step