Over the past few weeks, I had the opportunity to participate in the Google Gen AI APAC Academy – Cohort 3, where I worked through hands-on codelabs focused on building and deploying Generative AI agents on Google Cloud.
What made this experience valuable for me was that it wasn't just about experimenting with prompts or calling an LLM API. I got to work with AI agents, RAG, cloud deployment, data access, MCP, and production-oriented workflows.
I completed three tracks, each focusing on a different real-world use case.
☕ Track 1 — Customer-Facing Coffee Barista RAG Agent
The first project focused on building a customer-facing Coffee Barista agent using Google ADK and Cloud Run.
The goal was to create an agent capable of working with external knowledge rather than relying only on the model's existing knowledge.
The major concept I explored here was Retrieval-Augmented Generation (RAG).
Instead of simply asking the model to generate an answer, the system can retrieve relevant information and provide that context to the model before generating the response.
The overall idea can be represented as:
User
↓
AI Agent
↓
Retrieve Relevant Information
↓
Context + Model
↓
Generated Response
Working through this track helped me understand how retrieval can make AI applications more useful for domain-specific scenarios.
I also deployed the application using Cloud Run, giving me practical exposure to taking an AI application beyond the local development environment.
📊 Track 2 — Data Analytics Agent
The second track moved into a more data-oriented use case.
I built a Data Analytics Agent using Gemma 4, BigQuery MCP Server, and Direct VPC Egress.
This track was particularly interesting because the agent wasn't just generating text. It was designed around interacting with data.
One of the important concepts I explored was MCP (Model Context Protocol) and how it can provide a structured way for AI systems to interact with external tools and data sources.
The high-level workflow was:
User Question
↓
AI Agent
↓
MCP / BigQuery
↓
Data Retrieval
↓
Analysis
↓
Natural Language Response
This changed the way I think about AI applications.
An LLM by itself is powerful, but connecting it to reliable external data and tools makes it much more useful for real-world applications.
I also got exposure to Gemma 4 and cloud networking concepts through Direct VPC Egress.
☕ Track 3 — Coffee Shop Manager Productivity Assistant
The third track focused on building a Coffee Shop Manager Productivity Assistant and deploying it on Cloud Run.
This project helped me look at agents from a productivity perspective rather than just a conversational one.
The interesting part was thinking about how an AI system can assist with operational workflows and provide useful interactions for a specific user role.
This reinforced an important idea for me:
An AI agent becomes significantly more useful when it is designed around a specific workflow and given access to the right tools and context.
🧠 What I Learned
This academy gave me hands-on exposure to several concepts that I had previously understood mostly from theory.
- Agents are more than prompts
I started with the mindset of:
Prompt → LLM → Response
The projects showed me a broader architecture:
User
↓
Agent
↓
Reasoning / Decision
↓
Tools / Data / Retrieval
↓
Action
↓
Response
The ability to connect an AI model with tools, data, and application logic is what makes agent-based systems interesting.
- RAG is about grounding AI
The Coffee Barista project helped me understand why retrieval is important when an application needs to work with specific knowledge.
Instead of expecting the model to know everything, we can provide relevant information as context.
- MCP opens interesting possibilities
Working with the BigQuery MCP Server introduced me to another approach for connecting AI agents with external capabilities.
It made me think more about tool-enabled AI systems rather than isolated LLM applications.
- Deployment matters
Building something locally is only one part of the process.
Deploying the agents using Cloud Run gave me practical exposure to taking AI applications into a cloud environment.
- AI engineering is becoming a systems problem
The biggest takeaway for me was that building useful GenAI applications isn't only about choosing the right model.
It involves thinking about:
Models
Agents
Retrieval
Tools
Data
APIs
Networking
Deployment
Reliability
User experience
That's the direction I want to continue exploring.
🚀 What's Next?
This experience has strengthened my interest in building GenAI systems rather than just GenAI demos.
I'm particularly interested in exploring how agents can maintain context, use external tools, work with structured and unstructured data, and operate reliably as part of larger software systems.
The Google Gen AI APAC Academy gave me an opportunity to move from simply learning about these concepts to actually building with them.
A big thank you to Google Cloud and Hack2skill for creating this hands-on learning experience.
I'm excited to keep building, experimenting, and going deeper into Generative AI and AI systems engineering. 🚀
