Inspiration

We noticed that the most valuable insights from meetings, lectures, and brainstorming sessions usually disappear the moment the conversation ends. Notes are incomplete, recordings are too long to rewatch, and mind maps are almost always created after the fact — when the energy and context are already gone.

We asked ourselves: What if the mind map could grow while the conversation is still happening?

That question became EchoMap.

What it does

EchoMap listens to any conversation (in-person via microphone or online via browser) and builds a live, interactive knowledge graph in real time.

Key ideas, decisions, questions, and action items automatically appear as connected nodes.
You can click any node to expand it, ask the AI for clarification, or attach notes/files.
Multiple people can join the same live map and add their own thoughts simultaneously.
After the session, the map becomes a permanent, searchable knowledge asset that can be exported, shared, or continued later.

It turns fleeting conversations into structured, visual knowledge that actually lasts.

How we built it

Speech-to-text: Whisper (OpenAI) for accurate real-time transcription
Understanding the conversation: GPT-4o with a custom system prompt that extracts entities, relationships, decisions, and open questions
Live knowledge graph: React + React Flow for the interactive canvas
Real-time collaboration: WebSockets (Socket.io) so multiple users can edit the same map simultaneously
Backend: Node.js + Express
Storage: PostgreSQL for permanent maps + Redis for live session state

We also built a browser extension so EchoMap can listen to Zoom, Google Meet, and Microsoft Teams without any extra setup.

Challenges we ran into

Latency vs accuracy — Making the map update fast enough to feel magical while still being accurate was extremely difficult. We had to carefully balance chunk size, model choice, and post-processing.
Noisy conversations — Real meetings are messy (people talking over each other, side jokes, tangents). Teaching the AI what is actually important took many iterations of prompt engineering and filtering.
Real-time multiplayer graph editing — Preventing conflicts when multiple people drag, edit, or delete nodes at the same time required careful CRDT-like logic.
Keeping the map clean — Early versions produced chaotic, overcrowded graphs. We had to invent smart clustering and “importance scoring” so the map stays readable.

Accomplishments that we're proud of

The map feels genuinely alive — nodes appear and connect while people are still speaking.
Multiple users can collaboratively shape the same map in real time without breaking it.
The post-meeting experience is actually useful (not just another recording you never rewatch).
We got the end-to-end latency under 3 seconds on average.

What we learned

Real-time AI products live or die by latency and relevance filtering.
Conversations are much harder for AI than written text because of noise, context switching, and implicit meaning.
Collaborative interfaces need extremely careful state management.
The best features often come from watching real people get confused or frustrated during testing.

Most importantly, we learned that the future of note-taking isn’t better notes — it’s turning conversations themselves into living knowledge structures.