We Built an AI That "Daydreams": Our Google Cloud Hackathon Story

When our team first started the Agent Development Kit Hackathon with Google Cloud, we were driven by a single, nagging question: Why do today's powerful AIs feel so... passive? They are brilliant calculators, but they often feel like a "polite guest who won't speak until spoken to."

We wanted to build a true partner—an AI that could feel alive. One that could get curious with you, understand the flow of your thinking, and have an "inner life" of its own.

That idea became Wise. In just under two weeks, we built and deployed a working prototype of a system that we believe represents a new paradigm for human-AI interaction. This is the story of how we built it.

(Disclaimer: We created this project and article for the purposes of entering the Agent Development Kit Hackathon with Google Cloud. #adkhackathon)

The Vision: An AI That Thinks With You

Our goal was to build more than a chatbot; we wanted to build a co-creator. Wise is designed to be a calm, focused space for thinking, built on two core concepts:

1. A Proactive "Inner Life": Wise has a DaydreamAgent that 'thinks' in the background. As a proof-of-concept, it analyzes a dataset to find interesting anomalies and proactively greets you with a data-driven "spark." It solves the "passive AI" problem by being an active participant in the conversation.



2. 
   

   Cognitive Tuning: Wise doesn't have a single, static personality. By detecting the intent of your conversation, it switches between different Cognitive Lenses:

   
   
   
   
   
   • The Analytical Lens for data-driven research and tool use.
   
   
   • The Imaginative Lens for creative brainstorming and poetry.
   
   
   
   

How We Built It: The "SageMind Architecture"

Top AI researchers like Andrej Karpathy note that the future of AI applications lies in "multiple LLM orchestration." Our SageMind Architecture is our answer to this challenge.

For this hackathon, we made a crucial decision: instead of using a pre-built toolkit, we chose to build our own custom, lightweight agentic framework from first principles. This approach allowed us to demonstrate our deep understanding of agent orchestration by building the core logic from the ground up.

Our system is a unified Streamlit application where our custom-built Python agents are orchestrated to deliver a seamless experience.

1. The DaydreamAgent (The Proactive Dreamer):
   
   This agent is our Curiosity Engine. We used Python and Pandas to analyze a sample stock history CSV. The agent identifies the day with the highest trading volume and then uses Gemini 1.5 Flash to synthesize a natural language "spark" from that data point, which is then presented to the user.



2. 
   

   The ConversationalAgent (The Master Conductor):
   
   This is the main agent the user talks to. Our Streamlit front-end acts as the orchestrator, passing the active "Lens" (as a system prompt) and the conversation history to this agent. Based on user intent, this agent intelligently chooses which internal Tool to use:

   
   
   
   
   
   • In Analytical mode, it can use a "Data Analysis Tool" that leverages Gemini 1.5 Pro to write Python code on the fly, which is then executed to generate and display a Plotly graph.
   
   
   • In Imaginative mode, it uses its creative prompting to generate prose or poetry.
   
   
   
   

This custom architecture allowed us to build a sophisticated, decomposed workflow, a principle core to modern AI systems.

The Full Stack

• Application & Orchestration: Streamlit


• Agentic Core: Python, Pandas, Plotly


• AI Models: Google Gemini 1.5 Pro & Gemini 1.5 Flash


• Cloud Deployment: Docker, Google Cloud Run

What's Next? From Partner to Platform

This hackathon was just the beginning. The SageMind architecture is a powerful foundation. Our next steps are to build out the interactive Mind Map—the visual "second brain"—and expand our library of Cognitive Lenses.

Our long-term vision is to evolve Wise into a true Multimodal Conductor. This is a system that doesn't just use internal tools, but intelligently orchestrates a 'dream team' of best-in-class AI models and APIs from across the industry—connecting to Google BigQuery for enterprise data, using a web search tool for real-time information, and leveraging multimodal models to reason from images and documents.

Building Wise taught us that the future isn't just about bigger models, but about smarter, more empathetic architectures. We're excited to continue building that future.

Thanks for reading! You can check out our project on Devpost and see our final demo video [Here - Link to come].