🧠 MindStream: Agentic Mind Maps for Meeting Accessibility

Built at GemiHacks 2025

💡 Inspiration

Meetings are often overwhelming — especially for people with ADHD, cognitive disabilities, or those who are hard of hearing. Traditional summaries don’t capture structure, flow, or actionable points in a way that's easy to follow.

We built MindStream to make meetings more accessible by automatically generating mind maps from transcripts — giving users a visual, interactive way to understand what happened and what needs to be done.

👥 Who It's For

MindStream is designed for:

🔁 Neurodivergent users who benefit from visual organization of ideas
🧏 People who are hard of hearing, who may rely on transcripts
🧠 Anyone overwhelmed by long meetings or follow-up emails

Our goal: let users see the flow of ideas, understand structure, and retain action items without having to re-read a full transcript.

🧪 How It Works (AI Layer)

MindStream has two modes:

✅ Fast Mode

One-shot Gemini LLM call to extract mind map nodes from the full transcript.
Good for short transcripts, fast results.

🧠 Smart Agentic Mode (MCP-powered)

Gemini 1.5 is turned into an agentic system using Model Component Protocol (MCP)
Gemini is given three tools to call during reasoning:
1. extract_structure – extract nodes from transcript chunks
2. merge_maps – combine nodes into a full mind map
3. agent_memory – (optional) recall past sessions and build on them

These tools are accessed via FastAPI endpoints and JSON schemas, giving Gemini the power to reason, plan, and build high-quality maps from long or complex transcripts.

🛠️ How We Built It

Frontend: React + TypeScript + TailwindCSS + ReactFlow
Backend: FastAPI + MongoDB + Uvicorn
AI Engine:
- Gemini 1.5 (via google.generativeai)
- MCP-compatible tool interface
Agent: Gemini dynamically selects and calls tools as needed
Fallback: Classic LangChain-based agent for simpler queries

🔍 What It Looks Like

Drag-and-drop mind map
Visual branching of ideas and decisions
Highlights who said what and what needs to happen next

🧠 What We Learned

How to integrate Gemini 1.5 with real-time tool-calling (MCP)
The power of autonomous agents to handle long, messy inputs
How to design interfaces that support cognitive accessibility
How memory can help LLMs evolve their outputs across sessions

🚧 Challenges We Faced

Handling partial or malformed JSON from LLMs
Debugging tool-calling flow between Gemini and our backend
Managing long transcripts across multiple reasoning steps

💭 What's Next

Voice → transcript → mind map pipeline (fully automatic)
Save mind maps per user and let Gemini evolve them over time
Export to PDF, Notion, or calendar reminders
Real-time caption + mapping for live meetings

🏁 Final Thoughts

MindStream is more than a summary tool — it’s a visual accessibility layer for meetings. For people who struggle with attention, memory, or auditory processing, this project turns transcripts into structured, actionable visual maps powered by cutting-edge AI.