Agentiz Society

• 

  Dashboard

• 

  Live Map

• 

  Processes

Inspiration

Venture Capital analysts waste thousands of hours every year reading startup pitches that are scientifically incoherent or financially unviable. We realized that traditional LLMs are too eager to please and often hallucinate when asked to evaluate a pitch. We were inspired to build Agentiz Society, an objective, multi-agent orchestration platform where AI "experts" are forced to ruthlessly debate a startup's viability before any human investor wastes their time.

What it does

Agentiz Society is a fully autonomous AI Due Diligence platform. When a user submits a startup pitch, it kicks off a complex orchestration loop: Memory Retrieval: A Lead Investor agent queries a ChromaDB Vector Database to recall past investment memos and check for historical consistency. Task Delegation: The Lead assigns the pitch to a Tech Expert, Finance Analyst, and Risk Assessor. Autonomous Debate: The agents don't just chat—they execute recursive ReAct (Reasoning + Acting) loops to verify claims. They present their findings to the Lead Investor, who forces them to debate any conflicting views. Final Synthesis: The Lead Investor synthesizes the debate into a final Investment Memo (Invest or Pass) and permanently uploads the PDF/Markdown to an Alibaba Cloud OSS Bucket for immutable record-keeping.

How we built it

We built the frontend using React (Vite), focusing on a highly polished, glassmorphism UI featuring a live-updating "Agent Network Map" that visualizes the AI thought process in real-time. The backend is powered by FastAPI (Python). The core intelligence runs entirely on the Qwen-plus model via the Alibaba DashScope API. We engineered custom prompts and ReAct loops to keep the agents on track. For our RAG (Retrieval-Augmented Generation) memory, we integrated ChromaDB. Finally, we used the official Alibaba Cloud SDK to stream the final synthesized memos directly into an Alibaba OSS Bucket.

Challenges we ran into

Orchestrating multiple autonomous agents without them falling into infinite loops or agreeing too easily was a massive challenge. We had to heavily engineer the system prompts to ensure the Tech, Finance, and Risk agents maintained adversarial, highly critical personas during the debate phase. Additionally, we ran into a fascinating "memory poisoning" bug where our Vector Database accidentally saved an API Error as a past investment decision! We had to build robust error-handling safeguards to ensure only pristine, validated investment memos were written to our ChromaDB and Alibaba OSS storage.

Accomplishments that we're proud of

We are incredibly proud of bridging the gap between advanced backend AI engineering and beautiful frontend UX. Most multi-agent frameworks are restricted to boring command-line terminals. We managed to stream complex, multi-round AI debates directly into a stunning, consumer-ready interface. Successfully integrating Alibaba Cloud OSS and the Qwen API seamlessly into this workflow was the cherry on top.

What we learned

We learned that LLMs become exponentially more powerful when you restrict their roles and force them to debate. A single LLM asked to evaluate a pitch will often give a generic, balanced answer. But when you pit a specialized "Finance Agent" against a "Tech Agent" using the Qwen model, the resulting insights are astonishingly sharp and actionable.

What's next for Agentiz Society

In the future, we want to migrate our local vector memory over to Alibaba Cloud DashVector for enterprise-scale RAG. We also plan to expand the society by adding new specialized agents, such as a Legal Compliance Agent and an ESG (Environmental, Social, and Governance) Assessor, to provide even deeper due diligence.

Built With

• alibaba-cloud
• artificial-intelligence
• chromadb
• fastapi
• javascript
• machine-learning
• python
• qwen
• react
• vite