Skip to content
EpochC Start a project
· Afraz Khan

Case study: automating KYC to remove EUR 40,000 a year

Manual KYC is a tax that scales with your growth: every new customer is a person reading a document by hand. For a money-transfer platform (Ace Money Transfer), that manual onboarding was slow and expensive, and it could not keep up. Here is how we automated it end to end, and what it delivered.

The challenge

Verify customer identity across many document types, fast, without dropping the accuracy a compliance team requires. Off-the-shelf OCR could read text, but it could not reliably detect the document, extract the right fields, and know when it was unsure. And it had to run cost-effectively at volume.

What we built

A full-stack KYC verification system with three cooperating parts:

  1. Document detection and classification with a lightweight YOLOv8n model, trained on a custom-labeled dataset of 5,000+ real financial documents, and optimized for low-compute and edge deployment.
  2. Field extraction with PaddleOCR, using document layout rather than blind full-page OCR, so the system pulls the specific fields that matter per document type.
  3. Facial recognition with DeepFace, matching the ID photo against a live selfie to confirm the person matches the document.

Every extracted field carried a confidence score, so low-confidence cases routed to a human reviewer instead of silently passing, which is exactly the behavior a compliance team needs.

The results

  • EUR 40,000 per year in operating cost removed by automating onboarding that was previously manual.
  • 98% field-detection accuracy across multiple identity-document types.
  • 70% less compute than standard solutions, thanks to the lightweight model choice.

What made it work

  • A custom dataset. Generic models plateau. Labeling 5,000+ of the client’s real document types is what took accuracy from good to production-grade.
  • Low-compute by design. A lighter model at the edge beat a heavier one needing a GPU cluster, which mattered for both cost and data privacy.
  • Humans on the uncertain 2%. Confidence scoring caught exactly the cases that would otherwise have become compliance incidents.

This is what “production-grade” means in practice: not a demo that reads a clean passport, but a pipeline that handles the messy long tail and knows its own limits.

EpochC builds OCR and document-AI systems that ship. Start a project or hire on Upwork.