AUREA Collective Mind

Inspiration

Running a holding company with six business units from Valencia, Venezuela, I faced a daily reality: decisions get stuck in bottlenecks. A client request arrives, and it bounces between finance, legal, operations, and sales—sometimes for days. I needed a system that could think, negotiate, and execute like a C-suite team, not just answer chat prompts.

That is why I built AUREA Collective Mind: a society of AI agents where each one owns a domain, argues for its priorities, and reaches consensus to deliver end-to-end enterprise workflows without human gridlock.

What it does

AUREA Collective Mind is a multi-agent collaboration system built on Qwen Cloud. It deploys six specialized agents—Finance, Operations, Sales, Legal, Creator, and Strategy—that autonomously:

Decompose incoming business requests into sub-tasks.
Negotiate priorities and resource constraints in real time.
Resolve conflicts when agents disagree (e.g., Strategy wants to chase a lead, Finance flags budget risk).
Execute the final consensus: generating contracts, budgets, proposals, and project timelines.
Remember past interactions to improve future decisions across sessions.

A demo scenario: a prospect emails "We want AUREA to run our Madrid launch event." The Sales agent qualifies the intent, Strategy assigns priority, Finance builds a preliminary budget, Legal drafts the NDA, Operations checks resource availability against our December travel window, and Creator assembles the visual proposal. The human founder only intervenes at the final approval checkpoint.

How we built it

Orchestration layer: Python + FastAPI backend running on Alibaba Cloud ECS.
Agent framework: Microsoft AutoGen for multi-turn dialogue, role assignment, and conflict resolution between agents.
LLM engine: Qwen-Max for high-level reasoning and strategy; Qwen-Coder for implementation and tool-calling logic.
Memory: Redis for short-term session context; PostgreSQL (Alibaba Cloud RDS) for long-term persistent memory and preference storage.
Frontend: Streamlit dashboard to visualize agent conversations, decision trees, and final outputs.
Tooling: Each agent invokes external tools via APIs—calendar scheduling, document generation, budget calculators, and CRM connectors.
CI/CD: GitHub Actions for automated testing and Docker image builds before Alibaba Cloud deployment.

Challenges we faced

Conflict resolution: Getting agents to genuinely negotiate rather than blindly agree required careful prompt engineering and a "Devil's Advocate" arbitration protocol.
Context windows: With six agents conversing, token consumption exploded. We implemented hierarchical summarization so only critical memories survive long negotiations.
Tool reliability: External APIs fail. We built graceful degradation so an agent can request human intervention or retry with an alternative tool path.
Latency: Running on Alibaba Cloud from Venezuela meant optimizing ECS regions and caching strategies to keep response times under 3 seconds per agent turn.

What we learned

Multi-agent systems outperform single-agent pipelines on complex, ambiguous tasks—by 40% faster resolution time in our benchmark scenarios.
Qwen Cloud's tool-calling capabilities are robust enough to serve as the "brain" of an enterprise-grade agent society.
Memory is not just storage; it is the difference between a scripted bot and a learning organization. Forgetting obsolete data is as important as remembering key facts.

What's next

Migrate from Streamlit to a production Next.js frontend.
Add voice and WhatsApp interfaces for each agent.
Deploy the Madrid base instance in January 2027 to serve European clients with local latency.

Built With

alibaba-cloud-ecs
alibaba-cloud-rds
autogen
docker
fastapi
github
postgresql
python
qwen-cloud-(qwen-max
qwen-coder)
redis
streamlit

Submitted to

Global AI Hackathon Series with Qwen Cloud

Created by

I am the sole founder, architect, and developer of AUREA Collective Mind. I conceived the project from the ground up to solve a real bottleneck I face daily: running AUREA Enterprises, a six-company holding where decisions get stuck bouncing between finance, legal, operations, and sales.

My contribution spans the full stack:

Architecture & Design — I designed the multi-agent system with six specialized AI roles (Finance, Operations, Sales, Legal, Creator, Strategy) that collaborate, negotiate, and resolve conflicts autonomously. I defined the conflict arbitration protocol and the human-in-the-loop checkpoint.

Backend Development — I built the FastAPI backend in Python, integrated the AutoGen orchestrator for multi-turn agent dialogue, and connected the system to Qwen Cloud APIs (Qwen-Max for reasoning, Qwen-Coder for tool execution).

Cloud Infrastructure — I configured the deployment on Alibaba Cloud ECS with Docker, set up PostgreSQL RDS for persistent memory, and Redis for session state management.

Frontend & Visualization — I created the Streamlit dashboard that visualizes real-time agent conversations, decision trees, and final deliverables.

DevOps & Documentation — I wrote the complete README, architecture diagram, deployment scripts, and open-source license (MIT) for the public repository.

This was my first deep dive into multi-agent orchestration at enterprise scale. I learned that agent negotiation is harder than prompt engineering — getting agents to genuinely disagree and then reach consensus required careful prompt design and a "Devil's Advocate" arbitration layer. I also learned that memory management (what to remember, what to forget) is as critical as the agents themselves.

Every line of code, every architectural decision, and every pixel of the demo is mine. AUREA Collective Mind is not just a hackathon project — it is the operating system I will deploy inside AUREA Enterprises starting January 2027.

Geraldo Meneses

Updates

Geraldo Meneses started this project — Jun 02, 2026 05:30 PM EDT

Leave feedback in the comments!

Log in or sign up for Devpost to join the conversation.