National Identity Data Harmonisation

The problem

Operational challenge

Customer identity lived in two incompatible worlds: structured core-banking KYC on one side, unstructured biometric national ID records on the other. Names transliterated differently, dates disagreed, and duplicates were invisible — so verification was manual, slow and error-prone at branch level.

The approach

Implementation strategy

I designed a high-throughput harmonisation pipeline that normalises and fuzzy-matches record pairs, then escalates ambiguous cases to a CNN/Transformer-based biometric scorer. Confident matches resolve automatically; borderline cases surface in a review console with the evidence that produced the score, so an operator can decide in seconds instead of minutes.

Engineering trade-offs

Real-time banking constraints meant latency budgets came before absolute model accuracy. Rather than chase the last points of precision with a heavier model, I kept inference light and pushed uncertain cases to human review — the system is deliberately fast and honest about what it does not know.

What I built

Delivery highlights

• Built the ETL layer that reconciles transliteration variants, date formats and duplicate records between two independently-governed registries.
• Integrated a CNN-based face matching service with configurable thresholds, so risk teams can tune the automation/review split without a redeploy.
• Added a full audit trail over every automated decision to satisfy regulatory review.
• Containerised the inference services and put them behind CI/CD, cutting release cycles from days to minutes.

How it fits together

System architecture