A deployed document-ingestion system built with n8n and locally hosted vision-language models. It converts document images into structured records through multiple layers of extraction, validation and cross-checking. Each record remains linked to its source, while human-reviewed edits are protected from automated corrections.
Extracted fields pass through rule-based validation and a second-model review. Disagreements trigger OCR evidence checks and, where needed, a third model. The pipeline reconciles stored records, rechecks review locks before applying corrections and flags unresolved issues for inspection.
