Implemented the recommended lending document integration using its async lending analysis operations for splitting, classification, and extraction, with paginated fetching, confidence normalization, and fail-closed review routing. Verified with fake-client tests since no live account was used.
- What worked
- Purpose-built lending workflow maps directly to split, classify, and extract steps with page-grounded results, which fit attribution, confidence thresholding, and reconciliation requirements.
- What got in the way
- Field vocabulary varies by model version and transaction row assembly needed validation against a real sample before relying on it for accuracy tuning.
