The problem

In a meeting, everyone is taking notes, fact checking and chatting with an AI on the side, so nobody is really in the conversation.

The change

We did the opposite. Instead of giving people a better notes tool, we took the keyboard out of their hands. Nobody touches the board. Only Nemo does.

The project

Nemo is a voice AI facilitator for any team meeting. It turns what people say into a live shared board, with research and structure added as you talk.

Voice commands (examples)

• "Hey Nemo, today we're brainstorming how might we cut food waste in the cafeteria."
• "Nemo, merge 2 and 3", "rename 4 to smart scales", "move 5 and 7 to Pricing", "delete 6"
• "Nemo, look up how much food a canteen wastes per meal"
• "Nemo, explore pre-ordering systems in workplace canteens" (starts parallel sub-agents)
• "Nemo, let's do SCAMPER on 2", "six thinking hats", "a pre-mortem on 4"
• "Nemo, vote for 2, 6 and 10", "move on", "wrap up", "Nemo, stop"

How it works

Put one computer in the room. Anyone can follow the board on their own device with a 4 letter code. Ideas, concerns and questions become cards a few seconds after they are said. You steer Nemo by voice: merge, rename, group, link, vote, next phase, wrap up.

• NVIDIA Nemotron 3.5 Lightning on Nebius Token Factory is the fast brain. One tool call per sentence. Thinking is off on the voice path, since with it on the first token takes about 8.5 s.
• NVIDIA Nemotron 3 Super 120B on Nebius Token Factory is the coach. It spots stalls, clusters the board, asks challenge questions, runs the research sub-agents and writes the summary.
• NVIDIA Nemotron ASR Streaming NIM is the ears. Live transcription over a realtime WebSocket, with word boosting for "Nemo".
• NVIDIA Magpie TTS NIM is the voice, streamed chunk by chunk.
• Nebius AI Cloud hosts both NIMs on one L40S GPU VM, set up with our own scripts.
• Tavily is the real time web layer: fact checks and one search per research angle, called at runtime.
• TypeScript, Node 22, WebSockets, React Flow, elkjs, Vite, Tailwind v4, Docker.
• Hexagonal architecture: a pure domain core and swappable adapters for every provider.
• Single writer canvas: cards stay numbered, deduplicated and conflict free.
• Research runs in its own queues, so it never blocks the conversation.

Built With

• nebius
• nvidia
• tavily