Inspiration
GKE turning 10 inspired us to showcase how modern AI agents can augment legacy microservices without risky rewrites. In banking, three pains remain universal: clear financial guidance, real-time fraud detection, and responsive support. We aimed to prove that GKE + Gemini + MCP can add intelligence as sidecar capabilities, not invasive changes.
What it does
Adds three AI agents to Bank of Anthos without modifying core services: Financial Advisor: personalized insights, spending analysis, budget suggestions. Fraud Detection: anomaly detection blending heuristics, Isolation Forest, and Gemini scoring. Customer Support: natural language answers grounded in account, balance, and history. Uses an MCP server as the “API bridge” to existing microservices (balances, transactions, contacts). Runs on GKE with health checks, autoscaling, logs/metrics, and secret management.
How we built it
Base app: unmodified Bank of Anthos microservices. New components: FastAPI MCP server exposing read/write wrappers to existing APIs (JWT pass-through). Three FastAPI agents using Google Gemini (via google-generativeai) for reasoning. Fraud agent also uses scikit-learn (Isolation Forest) for structured anomaly signals. Packaging: Docker images stored in Artifact Registry; Kubernetes Deployments/Services; Secrets for Gemini API key. Ops: readiness/liveness probes, resource limits/requests, basic autoscaling setup. All deployed on GKE; no changes to the Bank of Anthos code paths.
Challenges we ran into
Passing and validating JWT across services while keeping the MCP stateless. Tuning request timeouts/retries so LLM calls don’t block critical paths. Normalizing data (transactions, timestamps) for hybrid ML + LLM scoring. Keeping a clean boundary to avoid “just one small change” in the core app. Ensuring reproducible builds and predictable costs during iteration.
Accomplishments that we're proud of
Non-invasive AI augmentation: zero edits to core microservices. A working trio of agents covering guidance, fraud, and support end-to-end. Robust MCP bridge that future-proofs additional agents and tools. Clear deploy story on GKE with observability and security basics in place. Comprehensive docs, diagrams, and a demo flow ready for judging.
What we learned
MCP is a powerful pattern to “snap” AI onto existing APIs safely. Grounding LLMs with real domain data (balances/history) boosts relevance. GKE autoscaling + resource governance are essential for AI workloads. Hybrid approaches (LLM + lightweight ML) improve fraud signal quality. Small operational details (probes, retries, secrets) make or break DX.
What's next for AI-Enhanced Bank of Anthos
Agent-to-Agent (A2A) orchestration for multi-step workflows (e.g., advisor collaborates with fraud). RAG for richer context (e.g., policy docs, product FAQs) and Vertex AI Search integration. Streaming processing for near real-time transaction scoring and proactive alerts. UI integrations in the frontend for first-class AI experiences and guardrails. Hardening: policy/guardrails, audit logging, multi-region GKE, FinOps tuning.
Log in or sign up for Devpost to join the conversation.