ThinkTank – A Framework for Generalizing Domain-Specific AI Agent Systems into Universal Collaborative Intelligence Platforms

The ThinkTank framework draws inspiration from the collaborative AI methodologies demonstrated in The Virtual Lab: AI Agents Design New SARS-CoV-2 Nanobodies with Experimental Validation, which showcased the effectiveness of multi-agent systems in scientific research and hypothesis generation.

The implementation of ThinkTank simulates a scientific team made up of local LLM-based agents enhanced with documents provided by the user, running fully offline via Ollama and any LLM models supported by Agno.

Please find the paper here: https://arxiv.org/abs/2506.02931

---

💡 What It Does

ThinkTank orchestrates collaborative research meetings between the following agents:

• Coordinator
  Synthesizes all ideas, resolves debates, and produces final meeting summaries and project decisions.

• Critical Thinker
  Points out flaws, inconsistencies, or gaps in reasoning, and demands scientific rigor.

• Domain Expert Agents
  Users can create any number of Domain Expert agents, each equipped with their own specialized collection in the vector store that serves as a RAG-based tool. Each expert agent is configured with distinct roles, areas of expertise, and unique contribution styles tailored to specific domains or organizational needs.

All meeting discussions and summaries are stored in a local memory database for future reference and summarization.

---

Requirements

1. Python version >= 3.11
2. git

---

Setup

1. Install Ollama and pull the model

brew install ollama               # Or follow https://ollama.com/download
ollama pull llama3.1              # This will download the model
ollama serve                      # Starts the Ollama server on localhost:11434

2. Clone and install Python dependencies

git clone https://github.com/taugroup/ThinkTank.git
cd ThinkTank

# Create a clean environment (recommended)
python -m venv .venv
source .venv/bin/activate

# Install dependencies
pip install -r requirements.txt

# To handle mermaid code from LLMs
npm install -g @mermaid-js/mermaid-cli

3. Run the script

streamlit run app.py

4. Meeting transcript can be downloaded as a Word(.docx) file from the app.

---

Citation

@article{surabhi2025thinktank,
  title={ThinkTank: A Framework for Generalizing Domain-Specific AI Agent Systems into Universal Collaborative Intelligence Platforms},
  author={Surabhi, Praneet Sai Madhu and Mudireddy, Dheeraj Reddy and Tao, Jian},
  journal={arXiv preprint arXiv:2506.02931},
  year={2025}
}

Team

• Devs: {Praneet Sai Madhu Surabhi, Dheeraj Mudireddy}, MS. Data Science '25, TAU Group
• Advisor: Prof. Dr. Jian Tao, Asst. Dir. of TAMIDS, Dir. of Digital Twin Lab