Case Study: Real-time Fraud Detection with Streaming Models

Client Overview

Client: Confidential Digital Bank
Industry: Neo Banking / Fintech
Duration: 3 Months (MVP to Deployment)

Business Challenge

The client was struggling to prevent fraud effectively due to reactive batch processing, high false positives, and inability to scale detection as UPI and card volumes grew. A modern fraud detection system was needed to detect anomalies in real time and reduce financial risk without increasing friction for genuine users.

Our Solution

Durga Analytics developed a high-performance streaming fraud detection platform combining machine learning, rules-based detection, and anomaly scoring. It was designed to operate at millisecond latency with hybrid signals across channels.

• Integrated real-time data ingestion using Apache Kafka
• Deployed ML classification model with Spark Structured Streaming
• Added rule engine for thresholds, velocity checks, geolocation flags
• Deployed containerized microservices with FastAPI
• Latency achieved: < 200ms per transaction

Key Outcomes

Technology Stack

Streaming & Data: Kafka, Spark, Redis
Modeling: Python, MLlib, Scikit-learn
APIs & Infra: FastAPI, Docker, PostgreSQL, Grafana
Deployment: AWS EC2, Lambda, EKS (pilot cluster)

Client Testimonial

“Durga Analytics delivered a real-time fraud detection platform that has significantly improved our operational control and reduced our exposure to high-risk behavior. The blend of ML and stream processing has changed the way we manage fraud.”
— VP, Fraud Risk Management

Next Steps

Durga Analytics is now supporting the client in extending the platform to include device fingerprinting, behavioral biometrics, and federated learning to combat new fraud vectors in the ecosystem.