Inspiration
While working with financial institutions, I noticed a critical gap: existing fraud detection systems require rebuilding entire applications or invasive code changes. When the GKE Turns 10 Hackathon challenged us to enhance microservices with AI agents, I realized we could solve this with an external intelligence network. What if we could add fraud detection to any application without touching a single line of existing code? SENTINEL was born from this vision - proving that distributed AI agents can observe, learn, and protect applications from the outside.
What it does
SENTINEL is an autonomous financial crime prevention network that enhances Bank of Anthos with real-time fraud detection capabilities. It deploys two specialized AI agents (SENTINEL-DETECT for behavioral analysis and SENTINEL-PREDICT for pattern recognition) that monitor transaction APIs and build a complete "shadow state" of user behavior through consensus-based detection. When suspicious activity is detected - like rapid card testing or unusual transaction patterns - the agents vote using confidence scores through our custom A2A protocol. The system runs completely external to the original banking application, requiring zero code modifications.
How we built it
Built on Google Kubernetes Engine with Gemini AI models powering each agent's analysis. The architecture centers on an MCP-like API wrapper that provides read-only access to Bank of Anthos transactions. Two specialized agents (DETECT and PREDICT) communicate through our custom A2A protocol, sharing insights and voting on fraud likelihood. Redis maintains the shadow state - reconstructing user behavioral patterns from API observations using our shadow_state_engine.py. Each agent runs in its own container with autoscaling, and Workload Identity provides secure access to Vertex AI. The consensus mechanism uses weighted voting to combine agent predictions.
Challenges we ran into
The biggest challenge was building consensus between two independent AI agents with different analysis approaches - how do you get multiple AI systems to agree on fraud detection? We solved this with our A2A protocol featuring weighted voting and confidence thresholds. Shadow state reconstruction was equally complex - rebuilding user behavior from limited API observations required sophisticated data modeling in our shadow state engine. Managing costs during development taught us important lessons about GKE resource management (learned the hard way with unexpected billing!). Getting the MCP-like wrapper to properly interface with Bank of Anthos APIs while maintaining performance took significant iteration.
Accomplishments that we're proud of
We successfully implemented a working multi-agent consensus system for fraud detection - a genuinely novel approach in the field. The shadow state reconstruction technique builds behavioral models purely from external API observations without requiring database access. Most importantly, we proved the concept works without modifying existing applications - SENTINEL can protect any microservice through external observation. The A2A protocol enables seamless agent communication and consensus voting. We're particularly proud that the system demonstrates real AI agent coordination on GKE using Gemini models.
What we learned
Distributed AI systems require different thinking than single-model approaches - consensus protocols become as important as the AI models themselves. We learned that Kubernetes is incredibly powerful for AI workloads when properly configured with Workload Identity and autoscaling. Managing cloud costs during development is crucial (definitely learned this lesson!). Building an MCP-like interface requires careful API design to maintain performance while providing comprehensive monitoring. Most importantly, we discovered that external AI observation can be more powerful than internal integration - sometimes the best way to add intelligence is from the outside looking in.
What's next for Sentinel
Immediate next steps: Implement the additional three agents (COMPLY, RESPOND, ANALYZE) that are architected but not yet built, expanding from our current two-agent system to the full five-agent network. Add comprehensive testing and validation to establish concrete accuracy metrics. Enhance the A2A protocol with more sophisticated consensus mechanisms. Long-term vision: Create a universal AI security layer that can protect any microservice application - making SENTINEL the standard for AI-augmented application security. We're also considering open-sourcing the A2A protocol as a standard for multi-agent communication in Kubernetes environments.
Log in or sign up for Devpost to join the conversation.