AegisFlow
Dual-mode multi-agent observability system for fraud detection and infrastructure monitoring
AegisFlow
A dual-mode multi-agent observability system for real-time fraud detection and infrastructure monitoring.
🎯 Built for Google Cloud Run Hackathon
AegisFlow demonstrates Agent Development Kit (ADK), Agent-to-Agent (A2A) communication, and production-ready serverless architecture on Google Cloud Platform.
🏗️ System Architecture

AegisFlow Architecture: Dual-mode system with real-time detection and interactive investigation
Background Worker Agents (Cloud Run)
- Agent 0: Log Parser - Processes heterogeneous logs from Bank of Anthos using rule-based parsing + Gemini LLM fallback
- Agent 2: Fraud Detector - Detects velocity attacks, brute force, large transactions using ADK agents
- Agent 3: Health Monitor - Monitors infrastructure health, detects service degradation and outages
A2A Investigation Agents (Cloud Run)
- Agent 4: Fraud Investigator - Investigates fraud patterns, queries BigQuery for historical alerts
- Agent 5: System Investigator - Analyzes infrastructure incidents and root causes
- Agent 6: Orchestrator - Routes user queries to appropriate investigator agents via A2A protocol
Data & APIs
- BigQuery - Central data warehouse for logs, alerts, and incidents
- Pub/Sub - Event streaming for agent communication
- API Service - REST API for querying BigQuery and GKE status
- GKE (Bank of Anthos) - Source application generating banking transactions and logs
✨ Key Features
🤖 Multi-Agent Collaboration
Agents work together using A2A protocol - orchestrator routes queries to specialized investigators
🔍 Real-Time Monitoring
Live dashboard showing fraud alerts, infrastructure incidents, and system health metrics
🧠 LLM-Powered Analysis
Uses Gemini for log classification, threat analysis, and root cause determination
💰 Cost Efficient
All services scale to zero when idle. Can disable GKE and still demo with cached data
🚀 How to Use
🏦 Try the Banking System Live!
Access Bank of Anthos to make transactions and trigger fraud detection in real-time
📖 Overview Tab (This Page)
Learn about the system architecture and features
💬 A2A Chat Tab
Ask natural language questions like:
- "What fraud alerts happened today?"
- "Show me incidents from the last week"
- "Which service has the most outages?"
📊 Agent Logs Tab
View real-time monitoring dashboard:
- GKE cluster status indicator
- Summary statistics across all agents
- Recent events, fraud alerts, and incidents
🎓 Technical Highlights
Google ADK (Agent Development Kit)
Agents 2 & 3 use ADK for structured agent definitions, BigQuery MCP toolsets, and automated tool calling for alert generation.
Agent-to-Agent (A2A) Protocol
Agents 4, 5, & 6 communicate via standardized A2A messages for inter-agent collaboration and orchestration.
Production-Ready Architecture
Separated concerns (workers vs APIs), event-driven design, centralized data in BigQuery, auto-scaling Cloud Run services, and comprehensive monitoring.
📦 Source Code
View on GitHub
Complete source code, deployment scripts, and documentation
💡 Demo Notes
Cost Optimization: The GKE cluster (Bank of Anthos) can be disabled to save costs (~$73/month). When disabled, the system continues to function with historical data from BigQuery, demonstrating the resilience and cost-efficiency of the architecture.
Check the 📊 Agent Logs tab to see the current GKE status!
Built with Google Cloud Platform • ADK • A2A • Gemini 2.0 Flash
Cloud Run • BigQuery • Pub/Sub • GKE