Introduction
As a Computer Science student at Vardhaman College of Engineering, I spend a lot of time working on backend systems. Having previously built a Retrieval-Augmented Generation (RAG) pipeline from scratch using Docker, Kubernetes, Jenkins, and Kafka during my software engineering internship at DevAI, I know how complex orchestrating AI can be.
That’s exactly why I joined the Google Cloud Gen AI Academy APAC Edition (Cohort 3). I wanted to see how enterprise-grade tools streamline the leap from local AI scripts to production-ready deployments. Over the past few weeks, I built a connected system of AI agents for a growing coffee shop business, deploying everything on Cloud Run. Here is a breakdown of what I built and the core architectural principles I learned along the way.
Track 1: Building a Customer-Facing AI Agent
The first challenge was building a personalized AI assistant for a coffee shop app that recommends orders to customers using the Google Agent Development Kit (ADK) and RAG.
The Build: Instead of manually writing and maintaining Dockerfiles like I have in the past, I learned how to use Google Cloud Buildpacks. Buildpacks automatically convert application source code into production-ready containers, drastically reducing manual maintenance.
The Concept: To make the chat experience truly interactive and handle continuous message exchanges, the application exposed a WebSocket endpoint rather than relying solely on standard HTTP requests.
Track 2: Turning Business Data into Strategic Decisions
As the coffee shop expanded nationally, we needed an agent to determine optimal new store locations—specifically ones friendly to cyclists.
The Build: I utilized Gemini alongside a BigQuery MCP Server to analyze public datasets for bike routes. We combined historical operational data retrieved via a tool with new user inputs to perform strategic analysis.
The Concept: A major takeaway here was the importance of the Principle of Least Privilege. We authenticated the application using a dedicated service account granted only the specific permissions needed to invoke the Gemini API and access the data, keeping the system highly secure.
Track 3: Automating Daily Operations
The final lab focused on the internal side of the business: deploying a personal productivity agent to help the shop owner check sales, inventory, and daily tasks by interacting with Google Sheets.
The Build: The architecture clearly separated the LLM's reasoning capabilities from its controlled interactions with external systems.
The Concept: Because this application executed commands and updated business records, security and governance were critical. I learned about running generated shell commands inside an isolated Cloud Run sandbox environment to minimize security risks. Furthermore, to implement "Human-in-the-loop" governance, the system required explicit user approval before the agent could apply modifications to shared operational records.
Final Thoughts Completing these three tracks completely shifted my perspective on AI architecture. It is no longer just about prompt engineering; it is about secure sandboxes, isolated environments, service accounts, and seamless Cloud Run deployments.
Now that I've graduated with my Gen AI Master Certificate, I'm gearing up for the Cohort 3 Ideathon, where I'll be building a secure, authenticated Personal Gemini Journal.