Real-time Fraud Detection & Personalisation
Scalable ML services for a regulated bank — batch and low-latency online inference powering fraud screening and personalised product recommendations for millions of customers.
- ~70%pipeline efficiency gain
- Batch + onlineinference in one design
Architecture · simplified
- Transactions & eventsStreaming and batch data
- FeaturesDistributed on Spark & Ray
- ModelsFraud scoring · recommendations
- ServingLow-latency APIs + batch jobs
- Banking systemsCustomer-facing & core banking
Context
The bank needed ML decisions — is this transaction fraudulent, what should this customer see next — delivered reliably into customer-facing and core banking systems, at petabyte data scale.
The challenge
Two very different latency profiles (real-time screening and large batch recommendation runs), strict regulatory expectations on auditability, and integration with systems that cannot go down.
What I did
- Designed scalable ML services supporting both batch and low-latency online inference on Spark, Ray and AWS.
- Built fraud-detection and anti-phishing capabilities for real-time transaction screening, alongside a product-recommendation framework for millions of customers.
- Integrated ML decisions into customer-facing and core banking systems through well-defined service interfaces and API contracts.
- Introduced CI/CD, observability and production-readiness practices so releases were auditable and reliable in a regulated environment.
Outcome
Pipeline efficiency improved by ~70%, and ML became a dependable, auditable part of production banking systems.