Kurodot AI: Multi-Agent

• 

  Multi-agent collaboration interface using color-coded sticky notes to assign tasks across different roles.

• 

  Editor agent

• 

  Project Manager (PM) successfully completing a workflow to transform a physical exhibition digital twin into a design.

• 

  VI Designer agent utilizing an external image as a style reference to guide the visual output.

• 

  Tech Producer agent

• 

  Data Analyst agent suggesting optimizations based on live traffic insights and user device trends.

• 

  High-density technical view of the Kurodot AI system architecture powered by Gemini and Google Cloud Run.

• 

  Cloud Run service metrics monitoring request counts and latencies to ensure stable agent performance.

• 

  Backend execution logs tracking the real-time orchestration and scheduling of the AI agent team.

"Meet Kurodot AI, your art agency team, powered by Gemini."

Project Overview — One sticky note activates an AI creative studio. Kurodot AI turns an exhibition URL into a fully curated banner through a team of agents including a PM, analyst, editor, and designer. Instead of a single prompt, the agents collaborate in a nonlinear workflow powered by Gemini.

Inspiration — I’m both a software lead and a digital artist. Real creative work doesn’t happen in a chat box — it happens in a studio with specialized roles working together. I wanted AI to work the same way.

What it does — Drop a sticky note and the workflow begins. The PM assigns tasks, the analyst reviews data, the editor writes the story, the designer creates visuals, and the tech producer exports the final assets. The agents also suggest improvements along the way.

How we built it — The system uses Gemini 2.5 Flash Image for multimodal generation and Gemini 2.0 Flash for reasoning. It runs on Python and FastAPI on Google Cloud Run, with Firestore for metadata and Cloud Storage for assets. The frontend uses lightweight JavaScript and WebSockets to show agents collaborating in real time.

Challenges we ran into — The biggest challenge was coordinating five agents working asynchronously. I built an "AgentCollaborationHub" to schedule tasks and keep a single source of truth while agents collaborate on the same canvas.

Accomplishments that we're proud of — The project moves beyond linear AI prompts into a collaborative multi-agent workflow. Each agent has a clear role and even provides proactive suggestions, like dark mode layouts or localization for different audiences.

What we learned — AI works best when it reduces coordination overhead instead of replacing human creativity. Gemini becomes far more powerful when embedded in a role-based agent system.

What's next for Kurodot AI — Next I want to expand the expertise library so users can plug in new agents for different industries. The long-term goal is a universal orchestration hub where complex ideas can be executed by a team of AI agents.

Built With

• cloud-build
• fastapi
• firestore
• gcs
• gemini-2.0-flash
• gemini-2.5-flash
• github
• google-cloud-run
• python
• umami-api
• vanilla-js
• vs-code
• websockets