Spend Guardian

Transaction summary
Flagged Transactions
Protected information
Logo

Inspiration

As a long-time user of GKE and practitioner on Google Cloud, alongside my passion for AI and enhancing user capability with AI, I’ve transformed Bank of Anthos to give it an agentic AI makeover called Spend Guardian. The idea was to take a classic microservices demo and show how AI agents can plug in seamlessly to deliver real, user-facing intelligence.

What it does

Spend Guardian is an agentic workflow layered on top of Bank of Anthos. It uses Gemini to: • Translate natural language questions into secure SQL queries over the ledger database. • Detect and alert users of unusual spending patterns in real time. • Let users ask questions like “How much have I paid Alice in the last 90 days?” and get clear, data-backed answers. • Act as a “guardian angel” for financial well-being by combining AI reasoning with strict guardrails.

How we built it

• Started from the Bank of Anthos microservices demo (running on GKE).
• Added an MCP Server that exposes Postgres via tool bindings, bridging natural language → SQL safely.
• Built the Spend Guardian service as a FastAPI gateway, integrating with Gemini models for tool use.
• Refactored the frontend to include a Spend Guardian chat box with optional recipient hints.
• Secured the deployment with Cloud Load Balancing, ManagedCertificates, Cloud DNS, and static IPs, making it secure to demo.

Challenges we ran into

• A lot of CrashLoopBackOff pods and image pull mismatches.
• Resolving broken pipes when bridging requests between services.
• Getting Gemini tool calls to properly translate to SQL queries instead of free-form answers.
• DNS + HTTPS provisioning delays while wiring a custom domain to the GKE ingress.

Accomplishments that we're proud of

• Took a static demo app and evolved it into a working AI agent-powered banking assistant.
• Successfully integrated Gemini function calling into a Kubernetes microservices app.
• Built a Bridge + MCP pattern that can be reused for other use cases
• Deployed a fully working AI-enhanced system with a custom domain + TLS

What we learned

• How to add tooling on top of the Bank of Anthos example
• How to translate user intent into structured tool calls while keeping AI outputs grounded in data.
• Kubernetes ingress + managed certs
• The value of agentic workflows in turning microservices into intelligent assistants.

What's next for Spend Guardian

• Add secure MCP servers with guardrails
• Expand to multi-agent workflows, like fraud detection, savings recommendations and investment workflows
• Explore fine-tuning on financial question datasets for improved accuracy, potentially including RAG