Taking enterprise invoice OCR from extraction to accountable QA_
An invoice-processing pipeline that combines OCR, field normalization, business rules, deduplication, and human QA instead of treating raw text extraction as the finish line.
- Invoice OCR
- PaddleOCR
- Document AI
- Human-in-the-loop
- Data Validation
Thousands of invoices arrived with changing layouts, weak scans, repeated documents, and fields that looked readable but were operationally wrong. Extraction coverage alone could not prove that the data was safe to send downstream.
Parallaxis built a staged document pipeline: image preparation, OCR, field mapping, rule-based fallbacks, confidence and mismatch checks, deduplication, reviewer feedback, and structured export.
The platform established full population of nine mandatory fields across a 2,438-row validation set while also exposing 545 QA mismatches for correction—turning apparent OCR success into measurable, reviewable data quality.
Overview
The client needed invoice data that could be used by finance and operational systems, not a text dump that merely looked correct on screen. The source documents included different supplier layouts, low-quality scans, inconsistent labels, and values that required business context to interpret.
Parallaxis built and tested an enterprise invoice-processing service around PaddleOCR and RapidOCR, with normalization, validation, fallbacks, deduplication, structured reports, and a QA loop for cases that could not be trusted automatically.
Why OCR accuracy was not the real target
A document can produce readable text and still fail the business process. A total may be captured from the wrong table row, an invoice number may be confused with a purchase order, or the same invoice may be processed twice. Field presence is therefore different from field correctness.
We defined the system around mandatory business fields and downstream constraints. Every extracted value needed a source, a normalized representation, and a reason to be accepted or sent for review.
The processing pipeline
Documents first pass through image preparation and OCR. Candidate fields are then identified from labels, layout, and document context before being mapped into a canonical invoice schema. Rule-based fallbacks handle recurring supplier patterns where a general extractor is uncertain.
The output layer applies type checks, date and amount normalization, cross-field comparisons, and a deterministic deduplication key. Each stage records enough context to reproduce why a value was selected.
Separating coverage from correctness
On the 2,438-row validation set, the system populated all nine mandatory fields. The QA comparison still identified 545 line-level mismatches. We retained both findings because hiding the second number behind a 100% fill rate would have produced a misleading quality claim.
Those mismatches became the improvement queue: recurring errors were converted into extraction rules or model adjustments, while ambiguous documents remained visible to reviewers. This is the practical human-in-the-loop layer that makes document automation accountable.
Structured output for downstream systems
Clean invoice records can be exported in a consistent schema and connected to internal applications through APIs. The service is deployed independently so OCR workers, validation logic, and consuming systems can scale or change without forcing a single large release.
Built for changing documents
Modern document parsers can convert page images and PDFs into structured JSON or Markdown, but enterprise invoices still require domain rules. The architecture keeps OCR, layout understanding, and business validation separate so a newer model can be adopted without discarding the accumulated knowledge of supplier formats and finance checks.
Project highlights
- PaddleOCR and RapidOCR extraction pipeline
- Nine-field canonical invoice schema
- Printed invoice and degraded-scan handling
- Layout-aware candidate selection
- Rule-based supplier fallbacks
- Date, amount, and identifier normalization
- Deterministic duplicate detection
- Mismatch reporting and reviewer feedback loop
- API-ready structured output
- Reproducible QA over 2,438 validation rows