• 

  architecture diagram

Project Story — FraudGuard AI

Inspiration

At our bank in Australia, we wanted to explore what the future of banking could look like. Traditional fraud alerts often notify customers after suspicious activity is detected, creating overhead and customer frustration.

We imagined a system where transactions could be verified instantly using biometrics, reducing operational overhead, minimizing losses, and improving customer experience. The idea of combining AI, biometrics, and explainability to proactively prevent fraud inspired FraudGuard AI.

What it does

FraudGuard AI is a multi-agent AI platform for banks that:

• Detects suspicious transactions in real-time.
  
• Validates biometric authenticity to prevent identity spoofing and deepfakes.
  
• Generates human-readable, explainable insights for CSRs.
  
• Integrates seamlessly into banking systems with serverless and containerized AWS infrastructure.
  

How we built it

• Frontend: React (Vite) for bank staff and system integrations.
  
• API Layer: Lambda functions or API Gateway for ingestion and orchestration.
  
• Backend: FastAPI running on ECS/Fargate.
  
• Agents:
  
  • Transaction Monitoring (XGBoost)
    
  • Deepfake Detection (SageMaker computer vision)
    
  • Evidence Collector Agent (Gathers User's Historical Transaction Data to study patterns)
  • Risk Assessment (Combines Insights of first 3 agents)
    
  • Escalation Handler Agent (Escalates the the fraudulent transaction reports)
    
• Storage & Infra: S3 (photos, models, logs), DynamoDB (transactions), SageMaker, Amazon Bedrock.
  
• Observability: CloudWatch, CloudTrail, X-Ray.

Challenges we ran into

• Data privacy & scarcity: Limited access to real banking data required synthetic datasets.
  
• Coordinating multiple AI agents: Ensuring agents worked together efficiently in real-time.
  
• Explainability: Translating complex model outputs into CSR-friendly, actionable insights.
  
• Performance: Balancing fast, real-time verification with compute-heavy models like Nova Pro.

Accomplishments that we're proud of

• Successfully integrated multi-agent AI orchestration for fraud detection.
  
• Built real-time biometric verification alongside transaction monitoring.
  
• Developed auditable, human-readable explanations for all flagged transactions.
  
• Achieved a scalable, serverless architecture using AWS Lambda, ECS, and API Gateway.
  

What we learned

• Multi-agent AI and LLMs can streamline complex decision-making in regulated industries.
  
• Explainable AI is essential for building trust with both CSRs and customers.
  
• AWS serverless + containerized architectures provide scalable, cost-efficient AI pipelines.
  
• Synthetic datasets and simulations are invaluable for testing AI in sensitive domains.

What's next for FraudGuard AI

• Implement a Feedback Loop Agent to continuously learn from CSR decisions.
  
• Add Drift Detection and Adaptive Learning to improve model performance over time.
  
• Develop a Human-in-the-loop UI for CSR review and real-time feedback.
  
• Expand external threat intelligence integration for enriched fraud signals.
  
• Explore fully proactive, real-time verification, aiming to prevent fraud before it happens.

Built With

• amazon-api-gateway-(api-routing-and-ingestion)
• amazon-bedrock-api-(llm-access
• amazon-bedrock-nova-pro-(llm-orchestration-and-explainability)
• amazon-dynamodb-(transaction-data-storage)
• amazon-ecs/fargate-(containerized-backend-runtime)
• amazon-s3-(storage-for-logs
• amazon-sagemaker-(deepfake-detection-and-ml-training)
• and-model-artifacts)
• bash-(automation-and-deployment-scripts)-frameworks-&-libraries:-fastapi-(backend-api-framework)
• boto3-(aws-sdk-for-python)-cloud-platforms-&-services:-aws-lambda-(serverless-orchestration)
• cloudtrail
• cloudwatch
• compliance)-databases:-dynamodb-(transaction-and-audit-logs)
• javascript/typescript-(frontend-with-react/vite)
• languages:-python-3.11-(backend-and-ai-logic)
• logs)-apis-&-integrations:-aws-sdk-/-boto3-(cloud-service-integration)
• models
• observability
• photos
• react/vite-(frontend-ui)
• rest-apis-(backend-agent-communication)
• s3-(unstructured-data:-images
• sagemaker-sdk-(deepfake-detection-and-model-deployment)
• shap-&-lime-(explainable-ai)
• x-ray-(monitoring
• xgboost-(fraud-detection-ml-model)