- Stage 1 (AI/LLMs Filtering): pre-filters transactions flagged for high-risk regions, large cash amounts, or non-standard accounts;
- Stage 2 (AI/LLMs Content Analysis/Multi-Agent): Agent deeply analyzes associated documents (internal communications, transaction memos) to determine if transactions share a Common Fraudulent Intent;
- Stage 3 (AI/LLMs Insight Generation/LangFuse): system automatically generates a comprehensive report detailing the reasoning and provides the LangFuse Audit ID for verification
flowchart TD
subgraph client_gateway_layer [Client and Gateway Layer]
A[User/Manager Input] --> B(Streamlit Frontend Proxy)
B -- Synchronous POST --> C{AWS API Gateway / Auth & Rate Limiting}
end
subgraph core_backend [Core Intelligent Backend]
C -- HTTPS/Auth --> D(FastAPI Backend: API Endpoint)
D -- Respond 202 Accepted --> C
D -- Async Trigger --> E(AWS SQS Queue: Agent Task Queue)
subgraph orchestration_exec [Orchestration and Execution]
E --> F(Agent Orchestrator Service - FastAPI/Uvicorn)
F -- Fails --> F_DLQ[AWS SQS: Dead Letter Queue]
F -- Health Check --> H(CloudWatch/Prometheus: System Metrics)
end
subgraph agent_stages [Sequential Execution within F]
F -- Stage 1 Filtering --> G1(PostgreSQL / Pinecone)
F -- Stage 2 Deep Review --> G2[External LLM Service with circuit breaker]
F --> I(Redis Cache: Rate Limit & Caching)
F --> J(LangFuse: Audit Trace Logging)
end
F -- Final Insight --> L(Webhook Callback / PostgreSQL)
L --> M(Streamlit: Updates UI Status)
end
subgraph data_observe [Data and Observability Layer]
H --> J
J --> H
end
style A fill:#DCE5F5,stroke:#333
style C fill:#FFE0B2,stroke:#FF9900
style D fill:#FFF4E8,stroke:#FFA000
style F fill:#E0F7FA,stroke:#00BCD4
style E fill:#F0F8E8,stroke:#00AA00
style F_DLQ fill:#FBEAEB,stroke:#D9534F
style I fill:#FFF4E8,stroke:#FF9900
style H fill:#E8F0FF,stroke:#3366FF