We automated KYC onboarding and removed €40,000 a year of manual review

The problem

Every new customer arrived as a photograph of an identity document, and every one of those photographs was read by a person. That queue set the pace of onboarding: applicants waited, reviewers burned hours on transcription, and the cost of the team scaled linearly with growth.

The documents themselves were the hard part. Passports, national IDs and driving licences from different countries share no layout, no field order and no fonts. Photographs arrived skewed, glare-blown, cropped and low-resolution. A generic OCR pass over that input produces confident nonsense, which in a compliance workflow is worse than producing nothing at all.

The team also needed the verification to prove that the person submitting the document was the person pictured on it — a separate problem from reading the text, and one with its own failure modes.

What we built

1. Classify the document before reading it

   A YOLOv8n classifier identifies the document type and locates it in the frame first. Knowing the layout up front turns an open-ended reading problem into a bounded extraction against a known template. We deliberately chose the nano variant so the model could run at the edge, which is where the 70% compute reduction against the baseline came from — the heavyweight pipeline never runs on frames that do not contain a usable document.

2. Normalise the image, then extract

   OpenCV handles deskewing, perspective correction, glare suppression and contrast normalisation before a character ever gets read. Most OCR accuracy is won or lost here rather than in the model: a correctly rectified image with even lighting is a far easier input than anything a phone camera produces unassisted.

3. Field-level extraction with PaddleOCR

   PaddleOCR reads the normalised crop, and extraction is scoped to the regions the classifier established. Fields are validated against the format each document type guarantees — checksum rules on document numbers, date ranges on expiry, character sets per issuing country — so a misread fails a validation rule instead of silently entering the record.

4. Match the face to the document

   DeepFace compares the portrait extracted from the document against the applicant’s submitted selfie, producing a similarity score rather than a boolean. The threshold is a policy decision the compliance team owns, not a constant buried in the code.

5. Route the uncertain cases to a human

   Every field carries a confidence score, and the pipeline is explicitly built so that low-confidence extractions escalate rather than pass. Automation removed the routine 90-odd percent of the queue; it did not remove the reviewer from the cases that genuinely need judgement. This is the design decision that made the system acceptable to compliance.

6. Serve it as an API

   The whole pipeline sits behind FastAPI, so onboarding calls one endpoint and receives structured, validated, confidence-scored fields plus a face-match result.

The outcome

• €40,000 a year in manual review cost removed, by taking the routine cases out of the queue entirely.
• 98% field-detection accuracy across mixed international document types and real-world photograph quality.
• 70% less compute than the baseline pipeline, by classifying cheaply before extracting expensively.
• Onboarding latency moved from human-queue time to API time, so applicants stopped waiting on reviewer availability.

Why this stack

PaddleOCR

Strong multilingual recognition and permissive licensing, which matters when documents span issuing countries.

YOLOv8n

The nano variant runs at the edge. Classifying before extracting is what cut compute by 70%.

DeepFace

Face verification with swappable backends, so the matching model can change without touching the pipeline.

OpenCV

Deskew, perspective correction and glare handling. Most accuracy is won in preprocessing.

FastAPI

Typed request and response models, so the contract with the onboarding service is enforced rather than documented.

Frequently asked