What's the best way to follow up a solo win at the Google Gemini 3 Seoul Hackathon? By diving even deeper into the Google AI ecosystem!
Just a few weeks ago, I participated in the hackathon as Team Mincasurong. Armed with Gemini’s incredible multimodal capabilities, I built GeminiSpace—a Vision-Language-Action (VLA) pipeline that generates spatial maps from photos—and was fortunate enough to take home 1st Place. However, as any engineer knows, building a rapid 7-hour prototype is very different from building a scalable, enterprise-grade service.
To take my current and future projects (like my upcoming multi-agent collaboration platform, Gouncil) to the next level, I needed to master the infrastructure that powers them. Recently, I attended the Google Cloud OnBoard - Google Skills Edition at the Westin Josun Parnas to learn exactly how to architect, govern, and deploy cloud-native Agentic services.
Here is a technical recap of the most impactful sessions and my takeaways as a robotics researcher transitioning into the "Agentic Era."
The morning kicked off with keynote sessions from Google Cloud and Google DeepMind leaders, showcasing the sheer scale of the AI-optimized stack.
It was amazing to see the expansion of the Vertex AI Model Garden, which now hosts over 200 models including the state-of-the-art Gemini 3, Gemma 3, and various open models. The demonstrations of GenMedia on Vertex AI—featuring models like Veo for video and Lyria for music—really highlighted the multimodal future we are heading toward. We are no longer just prompting text; we are orchestrating complex, multi-sensory data pipelines.
My main focus for the day was "Track 1: AI Build-up," which perfectly aligned with my goal of mastering multi-agent architectures. This track provided a deep dive into the modern AI developer's toolkit:
Agent Development Kit (ADK) & Protocols: We explored how to build multi-agent systems where agents don't compete, but seamlessly complement each other. Learning the clear distinction between the Model Context Protocol (MCP)—used to connect agents to tools and APIs—and the Agent2Agent (A2A) protocol for dynamic communication between distinct agents (like a Travel Agent collaborating with a Flight Agent) was a game-changer. This is the exact routing logic I need to make the Gouncil platform function.
Serverless AI on Cloud Run: Deploying these agents isn't just about writing code; it's about scalable infrastructure. We learned how to host frameworks like LangGraph and ADK directly on Cloud Run, seamlessly connecting them to Vector Databases (like AlloyDB or Cloud SQL), and building automated CI/CD pipelines with Cloud Build.
The absolute standout session for me was on Vertex AI-based Multi-layered Memory Architecture, presented by Generative AI Field Solution Architect Sang-woon So.
Handling long context and maintaining long-term memory is one of the biggest challenges in AI agents today. Vertex AI solves this elegantly by separating memory into distinct, human-like cognitive layers:
Session Memory (Short-term): Captures the contextual dialog state and tool return values during an active, ongoing session.
Personalized Memory Bank (Long-term): Stores user preferences, profiles, and accumulated insights across multiple interactions, enabling true personalization over time.
External Information Grounding: Accesses real-time data and location-based information via Google Search and Google Maps Grounding.
Internal Information Retrieval: Fetches context-specific internal company data using the RAG Engine and Google's highly efficient ScaNN-based Vertex AI Vector Search.
By orchestrating these layers, an AI agent can maintain context over long periods while seamlessly pulling in the exact right information from both the web and internal databases.
Beyond the technical deep dives, it was a fantastic day of networking, complete with a great lunch and some brilliant "Brew the Future" Google DeepMind swag.
I left the event with a clear understanding of the Vertex AI Agent Cloud environment and how to govern, orchestrate, and deploy these systems at an enterprise level. The gap between advanced AI software and physical-world robotics is closing fast. I definitely need more hands-on practice, but the blueprint is now clear. Time to fire up the IDE and start building the next generation of Agentic services! 💡☁️
#GoogleCloud #VertexAI #Gemini3 #ArtificialIntelligence #AIAgents #RAG #MachineLearning #GoogleCloudOnBoard #HackathonWinner #SoftwareEngineering #Robotics