NordPay Technologies, a Berlin-based B2B payments FinTech, needed to build an AI-powered transaction risk intelligence layer capable of processing 10M+ daily transactions with sub-100ms latency — while meeting Germany's stringent GDPR and BaFin regulatory requirements.
NordPay Technologies operates a B2B payment infrastructure platform serving mid-market German and Austrian enterprises. As their transaction volume scaled past 5M daily transactions, their legacy rules-based fraud detection system was generating 34% false positive rates — blocking legitimate payments and damaging client relationships. They needed to replace this with an AI-native risk intelligence layer, but their in-house engineering team lacked the ML expertise to build it.
Building a production-grade ML system for financial transaction scoring in Germany presented three simultaneous challenges: technical (sub-100ms inference at scale), regulatory (GDPR Article 22 requirements for automated decision-making, BaFin auditability mandates, and NIS2 cybersecurity obligations), and organizational (the system had to integrate with 6 existing microservices without downtime). No existing SaaS solution met all three requirements simultaneously.
We designed the solution around three non-negotiables: real-time inference performance, regulatory auditability, and zero-downtime integration.
A gradient boosting ensemble model trained on 18 months of transaction history, incorporating 147 features including merchant category, transaction velocity, device fingerprint, and behavioral patterns. Inference P99 latency: 68ms.
Every automated decision is accompanied by a ranked explanation of contributing factors — satisfying GDPR's right to explanation and BaFin's model governance requirements. Decisions are fully auditable with immutable logs.
Data minimization built into every pipeline stage. PII tokenization at ingestion, differential privacy for model training, and data residency controls ensuring all data remains within EU jurisdiction.
A Redis-based feature store pre-computes and caches transaction features, reducing inference latency from 340ms (batch query baseline) to under 70ms at the 99th percentile.
We ran the new model in shadow mode alongside the legacy system for 6 weeks, comparing decisions on live traffic — achieving 94% precision before hard cutover with zero client impact.
Real-time monitoring for the fraud operations team — live score distributions, model drift detection, manual review queues, and regulatory reporting exports.
The Zyrova team understood something most AI vendors miss: in German financial services, auditability is not a feature request — it is a commercial prerequisite. They built explainability into the model architecture from day one, which made our BaFin approval process straightforward.
Dr. Henrik Bauer Chief Technology Officer · NordPay Technologies