The Cognitive Swarm

A real-time, multimodal brainstorming application that enables large groups to collaboratively generate, cluster, and converge on ideas through voice, voting, and 3D visualization.

Project Lead: Keshav Dalmia

3D Swarm Visualization
Live Voice Anchor
QV Quadratic Voting

How It Works

A session moves through three admin-controlled phases: Explore, Vote, Forge

0

Create or Join a Room

An admin creates a room with a brainstorming topic and gets a 6-character room code. Participants join with the code and a display name. Admin and participant roles are distinct—the admin transitions between phases when the group is ready.

1

Explore

Participants speak ideas aloud while the AI anchor listens, responds, and nudges quiet voices. Tool calls extract discrete ideas from natural speech and place them in the 3D swarm by semantic similarity. Three background agents work in parallel: a synthesizer discovers connections, a devil’s advocate challenges weak ideas, and a direction suggester prompts new angles when conversation stalls.

2

Vote

The admin transitions the room to the Vote phase. Each participant gets a fixed pool of credits and allocates them across ideas using quadratic voting—each additional vote on the same idea costs more, surfacing genuine group preference rather than the loudest voice.

3

Forge

The admin transitions to Forge, and the top-weighted ideas are synthesized into a Mermaid diagram—flowchart, mindmap, ER diagram, class diagram, or journey map, with the type inferred from the topic. The artifact is shareable and exportable.

Technical Architecture

Production-deployed on GCP with full IaC and CI/CD

Real-Time Sync

  • Socket.IO with Redis adapter for horizontal scaling
  • Shared room state across all participants
  • 15-second admin reconnect grace before room closes
  • Sub-second latency for smooth collaboration
🎤

Gemini Live Anchor

  • Full-duplex, interruption-aware voice via Gemini Live
  • ~1 JPEG/sec camera stream alongside audio
  • Tool calling extracts ideas directly from speech
  • Anchor praises contributors and nudges quiet voices
🧠

Semantic 3D Swarm

  • Gemini embeddings drive 3D position by similarity
  • Clusters form visible constellations, not just colors
  • 200-idea cap with lowest-weight auto-prune
  • Synthesizer agent draws edges between related ideas
🤡

Background Agents

  • Synthesizer—discovers connections between ideas
  • Devil’s Advocate—challenges weak reasoning
  • Direction Suggester—prompts new angles when stalled
  • All three run in parallel during the Explore phase
⚖️

Convergence & Forge

  • Quadratic voting—each extra vote costs more
  • Credit allocation prevents dominant voices
  • Topic-aware Mermaid synthesis: flowchart, mindmap, ER, class, journey
  • Diagram type inferred from the room topic
☁️

Production Infrastructure

  • GCP Cloud Run, Memorystore Redis, Cloud Firestore
  • Secret Manager + Artifact Registry + VPC Connector
  • All resources defined in Terraform
  • GitHub Actions for build, deploy, and Terraform apply

Use Cases

Designed for any scenario where groups need to think together

Classroom Brainstorms

Enable large lecture halls to collaboratively generate and prioritize ideas in real time, giving every student a voice through quadratic voting.

Design Thinking

Support ideation and convergence phases of design sprints with AI-powered clustering that surfaces themes humans might miss.

Strategic Planning

Help teams move from divergent brainstorming to structured action plans with automatically generated diagrams and artifact synthesis.

Research Workshops

Facilitate collaborative knowledge synthesis where participants contribute through voice, see ideas cluster in 3D, and vote on research priorities.