SentinelStream: AI-Driven Cross-Border Fraud Prevention

A serverless fraud detection engine using Amazon Nova 2 Lite and Java 21 to identify high-risk e-commerce transactions in real-time with automated SNS alerts and DynamoDB persistence.

Comment

Inspiration

Global e-commerce faces increasingly sophisticated cross-border fraud. I was inspired to build a solution that bridges the gap between high-speed transaction processing and intelligent reasoning. By combining the ultra-low latency of Amazon Nova 2 Lite with a robust AWS Serverless architecture, I aimed to provide merchants with a real-time "AI Shield" against financial threats.

What it does

SentinelStream is an automated fraud prevention engine. It ingests global order data (amount, currency, IP, location) and leverages Amazon Nova 2 Lite to perform real-time risk assessment. High-risk transactions (Score > 0.8) trigger instant SNS email alerts, while all AI-generated insights and risk scores are persisted in Amazon DynamoDB for comprehensive security auditing and historical analysis.

How we built it

The system is built on a 100% Event-Driven Serverless architecture. ●Backend: Developed with Java 21, leveraging modern performance features and AWS Lambda Core libraries for minimal cold-start latency. ●Intelligence: Powered by Amazon Bedrock (Amazon Nova 2 Lite) for high-speed, cost-effective fraud inference with advanced reasoning capabilities. ●Data & Messaging: Integrated Amazon DynamoDB for scalable persistence and Amazon SNS for real-time incident response. ●Infrastructure as Code: Managed and deployed via AWS SAM, ensuring a reproducible and secure cloud environment.

Challenges we ran into

One major challenge was navigating AI Safety Filters. When the model encountered highly anomalous or "suspicious" test data, it occasionally triggered safety guardrails. I implemented a robust defensive programming logic to intercept these blocks, ensuring the system intelligently flags them as high-risk (0.99) instead of failing, thus maintaining system reliability under extreme scenarios.

Accomplishments that we're proud of

I am proud of successfully migrating the entire intelligence core from Claude to Amazon Nova 2 Lite during the development phase. This optimization achieved a superior speed-to-cost ratio while maintaining high inference accuracy. Achieving a fully functional, production-ready serverless workflow in such a short time was a significant milestone.

What we learned

I gained deep insights into Prompt Engineering for Amazon Nova 2, specifically in enforcing strict JSON schemas and managing model guardrails. I also mastered the art of optimizing Java 21 for Lambda environments (using the Maven Shade Plugin for Uber-Jars) and learned how to orchestrate multi-stage AWS services into a cohesive, secure Agentic AI system.

What's next for SentinelStream

The roadmap includes implementing Advanced Behavioral Analytics to detect bot patterns and expanding the notification system to Slack/Teams. I also plan to explore Amazon Nova 2 Pro for multi-modal analysis of transaction metadata and develop a React-based Dashboard to visualize real-time fraud trends from DynamoDB.

Built With

Share this project:

Updates

Submission history