HomeSelected engagements

02 / 05Case study

KYC automation pipeline

Kuda Technologies · African neobank, 7M+ customers

We cut KYC from 7 days to 30 min using VLMs. The hard part wasn’t the model.

7 days → 30 min

verification time

KYC automation · Kuda Technologies

7 days → 30 min

median verification time per submission, end to end

Routing by confidence. Schematic. Above the threshold τ the model decides; below it, a person does. Measured: under 2% of submissions go to a person, 98% of automated decisions correct, human review down from 176 to 2 hours a month. 4 h at launch, 30 min median after iteration.
Built with Gemini Flash, Qwen3 8B VL, Prompt engineering Write‑up June 2026 · 9 min read

01

The problem.

KYC verification took 7 days end to end, requiring a large manual review team. A hard bottleneck to customer onboarding at scale.

“Eight customer experience workers spent roughly an hour per day each manually reviewing KYC documents. Total: about 176 hours a month. Each submission took 7 days end to end. The review queue never emptied.”

“The Central Bank of Nigeria had updated its KYC regulations to require physical address verification for every customer. The manual process was the bottleneck.”

An address on a utility bill

“Plot 5, Adeola Odeku, beside the blue building, behind the church.”

“Landmarks and relative locations replace street numbers. The address fields those tools expect do not exist in the same way here.”

“This is not a model accuracy problem. It is a data standardization problem that most ML content assumes away.”

From the write‑up: The Problem That Benchmark Scores Don't Capture

02

What I built.

Scoped, designed, and shipped an end to end KYC automation system using vision language models (Gemini Flash, Qwen3 8B VL), taking it from initial problem definition through model selection, fine tuning, and production deployment. System classifies identity documents, detects fraudulent submissions, and auto approves customers exceeding confidence thresholds.

02.1

Model Selection Under Constraints

“The deadline was one week. CBN compliance urgency. Not enough time for model benchmarking.”

“I chose Gemini Flash because it was available immediately, had strong OCR and document understanding, and the latency profile suited batch processing. Built the initial system in one week. It processed 400K+ submissions.”

“I later expanded to Qwen3 8B VL for sensitive and edge cases that required self hosted inference. Fine tuned with SFT LoRA on 10K curated records.”

From the write‑up

1 week to the initial system

02.2

The Consistency Problem

“On documents with descriptive addresses (landmarks, relative locations, area names), reviewers agreed on the correct extraction roughly 75% of the time.”

“The system produced identical extractions on the same inputs 100% of the time.”

“The system was not better because the model was smarter. It was better because it was consistent. Consistency is what compliance actually needs.”

From the write‑up

75% reviewer agreement on descriptive addresses

02.3

The Evaluation Harness

“Early in development, the model showed 96% aggregate accuracy on extraction. That number looked fine. When we split by document type, accuracy on utility bills was 99% but accuracy on the category we had labeled ‘photographs’ was 83%.”

“After that, we never reported an unstratified accuracy number again.”

From the write‑up

83% accuracy on “photographs”, inside a 96% aggregate

02.4

The Regression Problem Nobody Talks About

“Adding a new function: a new document type, a new address format variant could silently break existing behavior.”

“The new format extracted correctly. But the model's confidence calibration shifted on an unrelated existing document type, causing roughly 1,200 submissions that would have been auto approved to flag for manual review instead. The evaluation harness caught it because it ran the full test suite on every model update, not just the new function.”

From the write‑up

~1,200 submissions a silent regression would have sent to manual review

02.5

The Pipeline Architecture

“Fraud detection lived inside the VLM extraction step, not as a separate classifier.”

“Baking it into the extraction step means the same model that reads the document also assesses its authenticity.”

“The simpler architecture was the better value.”

From the write‑up · See the schematic

03

Architecture.

Schematic

  1. 01 KYC submission 80K+ customers a month
  2. 02 Document classification utility bill · bank statement · photograph
  3. 03 Field extraction and fraud detection (model step) one VLM step · Gemini Flash, Qwen3 8B VL
  4. 04 Confidence scoring can the decision be automated?
  5. 05 Routing
    • Approval
    • Decline
    • CS review, under 2%

Evaluation harness · full suite on every model update

  • Document type
  • Extraction field
  • Quality tier
  • Format variant

Feeds back into step 03

Schematic of the document flow described in the write-up.

04

Results.

  • 7 days → 30 min

    median per submission, end to end

  • 4 h

    at launch, before iteration

  • 176 h → 2 h

    human review per month

  • 98%

    automated decision accuracy

  • 400K+

    submissions processed

  • System now handles the majority of KYC decisions autonomously, with human review reserved for edge cases only.

05

Lessons.

“The initial system was built for approvals only, not declines. This was a deliberate risk decision, not a technical limitation.”

From the write‑up: What Didn't Work

“The model was running in a week. The infrastructure around it: evaluation, rejection handling, confidence calibration, compliance documentation took months. That is not a weakness. That is the point.”

From the write‑up: What Didn't Work

“If you are deciding between labeling 10K examples yourself and pulling 50K from an API, label the 10K.”

From the write‑up: The Curation Insight