Bank of Anthos AI Agents

• 

  Bank of Anthos UI Login

• 

  Bank of Anthos Dashboard (account view)

• 

  kubectl get deploy (BoA core + agents running)

• 

  kubectl get pods (adk-gateway pod running + describe env)

• 

  Step-by-step rollout script / sanity checks

• 

  Swagger UI (API docs for /fraud/score)

• 

  Swagger Response Schema - Vertex used case

• 

  Fraud Agent curl demo (txn-allow)

• 

  Fraud Agent curl demo (txn-review)

• 

  MCP server deployment + IAM binding

Inspiration

Bank of Anthos is a classic microservices demo, but it lacked intelligence in decision-making. Inspired by real-world banking needs—fraud prevention, credit scoring, and compliance—we wanted to supercharge it with agentic AI while leaving the core services untouched.

What it does

We added new AI-powered agents running on GKE that interact with Bank of Anthos through existing APIs: • Fraud Sentinel Agent → detects suspicious transactions in real time. • Creditworthiness Co-Pilot → AI-assisted scoring for loan applications. • Compliance Agent → monitors transactions for policy or regulatory breaches.

All powered by Gemini AI (via AI Studio) for natural reasoning, with safe fallback heuristics.

How we built it

• Deployed agents as containerized FastAPI services in a separate Kubernetes namespace.
• Used Model Context Protocol (MCP) to wrap Bank of Anthos APIs (user profile, transaction history).
• Integrated Gemini AI via AI Studio with API key authentication.
• Added safe fallbacks so the system continues working even if AI is unavailable.
• Delivered infra with GKE, Artifact Registry, and automated rollouts.

Challenges we ran into

• Ensuring AI agents didn’t require touching the existing Bank of Anthos core code.
• Debugging multi-arch Docker builds on a Mac M1 dev machine with Colima + buildx.
• Handling differences between AI Studio (API key) and Vertex AI (service account).
• Getting structured, explainable “reasons” out of the model consistently.

Accomplishments that we're proud of

• First working version of Bank of Anthos enhanced with live agentic AI calls.
• Clean separation: AI logic in agents, core microservices untouched.
• Demonstrated end-to-end flow: curl → fraud agent → MCP → Bank of Anthos APIs → Gemini decision.
• Achieved reproducible builds and pinned images with digests in GKE.

What we learned

• How to extend production-grade microservices safely with external AI agents.
• The importance of fallback strategies for reliability in AI-enhanced apps.
• How MCP standardizes AI-agent interaction with microservices.
• Practical tradeoffs between Vertex AI and AI Studio integration.

What's next for Bank of Anthos AI Agents

• Multi-agent orchestration (A2A): Fraud signals influencing credit decisions.
• kubectl-ai integration: Natural language ops for GKE management.
• Stronger explainability: richer audit logs for compliance.
• Scaling to more domains: extending beyond fraud/risk into personalized banking.

Built With

• agent2agent(a2a)
• artifactregistry
• colima(macm1-builds)
• dockerbuildx
• googlecloud-iam
• googlegemini(ai-studioapikey-integration)
• googlegenai-sdk
• googlekubernetesengine(gkeautopilot)
• java(bankofanthosmicroservices)
• kubectl-ai-forsafeops
• modelcontextprotocol(mcp)
• python(fastapi)
• restapis