Gemini FraudShield

An intelligent fraud detection system that acts as an external "smart brain" for Bank of Anthos, using AI agents (Gemini AI) to provide real-time fraud analysis with explainable AI reasoning.

Comment

Inspiration

The inspiration for this project came from the critical need for real-time fraud detection in modern banking systems. With the rise of digital transactions and sophisticated fraud schemes, traditional rule-based systems are no longer sufficient. We were inspired by:

  • Google’s Gemini AI capabilities for natural language reasoning and pattern recognition
  • Bank of Anthos as a realistic microservices banking platform
  • The challenge of external monitoring without modifying core banking infrastructure
  • Real-time AI analysis that provides explainable fraud decisions

What it does

  • Monitors every Bank of Anthos transaction in real time without changing core services
  • Analyzes transactions using Gemini AI and returns fraud score, risk level, recommendation, and human-readable reasoning
  • Displays results in a professional, Bank of Anthos–styled dashboard with filtering and auto-refresh
  • Supports transaction simulation for demos/tests and a demo control panel to clear data and control load generation
  • Handles AI rate limits gracefully with a high-quality fallback analysis so the system remains usable under quota pressure

How we built it

Non-invasive interception pattern:

  • Redirected TRANSACTIONS_API_ADDR via Kubernetes ConfigMap to a Transaction Interceptor proxy
  • Interceptor forwards to ledgerwriter and asynchronously submits transactions to Fraud API

Fraud Analysis service:

  • FastAPI backend integrates Gemini 1.5 Flash for natural-language reasoning
  • Risk scoring and explanations; attaches analysis back to transaction records for the dashboard
  • Added endpoints for transactions, alerts, health, and demo reset (clear-transactions)

Dashboard:

  • Flask UI using Bootstrap and Bank of Anthos styles
  • Real-time updates (auto-refresh), account filtering (from/to), and simulation form

Demo control:

  • Lightweight controller to start/stop loadgenerator and clear API data to manage quota
  • Deployed on GKE with LoadBalancer services for external access; logs and health checks for verification

Challenges we ran into

Service Integration Complexity

Challenge: Integrating with Bank of Anthos without modifying core code
Solution: Used Kubernetes ConfigMap redirection and proxy patterns

AI Rate Limiting

Challenge: Gemini API quota limits during high transaction volumes
Solution: Implemented intelligent fallback analysis with detailed reasoning

Real-time Performance

Challenge: Maintaining sub-second response times while performing AI analysis
Solution: Asynchronous processing with immediate acknowledgment

Template Rendering Issues

Challenge: Jinja2 template errors when transaction data was incomplete
Solution: Defensive programming with graceful degradation

Demo Environment Management

Challenge: Controlling transaction generation for optimal demo experience
Solution: Built comprehensive demo control system with transaction clearing

Accomplishments that we're proud of

  • Achieved seamless zero-modification integration with Bank of Anthos
  • Delivered explainable AI fraud decisions that are understandable to operators
  • Built a clean, Bank of Anthos–consistent UI with filtering and auto-refresh
  • Ensured resilience under AI rate limits with robust fallback analysis
  • Added demo-friendly controls: one-click data reset and load generation control
  • Attached AI analysis directly to stored transactions for a complete audit trail

What we learned

Throughout this hackathon, we gained deep insights into:

  • Designing non-invasive integrations for microservices with proxy/redirection patterns
  • Using Kubernetes ConfigMaps and service discovery to transparently adjust traffic flow
  • Balancing AI power with reliability by implementing robust fallbacks
  • Applying async patterns to keep UX snappy while heavy analysis runs in the background
  • Hardening production-facing dashboards with defensive templating and graceful degradation

Next steps for Fraud Shield

  • Microservices Architecture: Understanding how to integrate external services with existing microservices without disrupting core functionality
  • Kubernetes Service Mesh: Leveraging ConfigMaps and service discovery for transparent traffic interception
  • AI Integration Patterns: Implementing robust fallback mechanisms when AI services hit rate limits
  • Real-time Processing: Balancing immediate response times with comprehensive AI analysis
  • Production-Ready Monitoring: Building systems that handle edge cases gracefully

Built With

Share this project:

Updates