GreenKnip Document AI Onboarding: From PDF Reports to Structured Metrics
A .NET document automation service for vendor onboarding, OCR cleanup, Dutch annual-report parsing, validation, and structured financial metrics.
- YEAR
- CATEGORY
- Document AI .NET OCR FinTech OpenAI
- BY
- Aleksei Skorjak
- FILE
- greenknip-document-ai-onboarding.prj
A production onboarding service that turns vendor financial documents into validated fields instead of another manual review queue.
The problem: annual reports are messy input
Vendor onboarding depends on documents that rarely behave. PDFs arrive with scanned pages, uneven tables, mixed Dutch financial terms, page headers, footers, and values that look similar but mean different things.
The first version used a Python OCR and NLP pipeline with PaddleOCR and LayoutLM. It proved the direction, but extraction consistency was not good enough for production onboarding.
The production approach
I rebuilt the service around .NET and C# microservices. The pipeline ingests PDFs, runs OCR, cleans noisy text, normalizes fields, asks OpenAI models to parse unstructured annual-report language, then validates the structured output before it reaches onboarding users.
- PDF ingestion and text extraction service.
- Tesseract OCR tuned for the document set.
- Layout-aware cleanup for headers, tables, repeated labels, and noisy text.
- OpenAI parsing step for Dutch financial language and metric extraction.
- Validation logic for common annual-report formatting issues.
- Test coverage around known failure cases.
Why simpler beat heavier
The useful move was not adding more model layers. It was making each stage easier to observe and easier to test. OCR had a clear job. Cleanup had a clear job. Parsing had a clear job. Validation caught the cases that looked plausible but were wrong.
That structure made failures easier to debug than a single opaque pipeline.
Result
The service reduced processing time by about 50% and improved reliability for vendor onboarding.
Stack: C#, .NET, REST APIs, microservices, Tesseract OCR, OpenAI GPT-4, document parsing, data validation, FinTech workflows, regulatory document handling.