A Reactive Fraud Monitoring Engine for Instant Payments
summary
Reactive pipeline takes channel events, fans them into parallel predictive modules and aggregates partial scores into an instruction: allow / stop / step‑up. Implementation uses Apache Kafka as the event backbone, Apache Flink and microservices for processing, with many legacy Python models deployed in Docker. Targets are sub-350 ms response and handling ~1.2M transactions/day with peaks of 200–500 tx/s; 60–70% of transactions are whitelisted to skip scoring. Challenges: limited elasticity and DevOps complexity from non‑JVM Dockerized models; staging/data‑lake replay for model validation.
AI-written summary · about
chapters (9)
- 0:00Introduction to reactive fraud monitoring
- 2:45Legacy fraud detection challenges
- 4:22Instant payments and new risks
- 6:22Reactive fraud monitoring engine
- 8:45Architecture design and scalability
- 13:00Implementation and model integration
- 15:34Model validation and testing
- 17:24System performance and limitations
- 19:56Future improvements and Q&A
related talks
25mReal-Time Fraud Detection at Scale
51mReal-time Fraud Detection: Challenges and Solutions
27mArchitecting Sub-Second Analytics Pipelines
47mReal-time Fraud Detection Challenges and Solutions
13mTransforming Financial Services with Real-Time Analytics
37mThe Role of Machine Learning in Today’s Payment Ecosystems
18mKubernetes to serve a fraud detection model
29mBuilding Resilient AI Platforms: Financial Fraud Detection