🎙️ Helios - AI-Powered Conversation Knowledge Graphs

💡 Inspiration

Ever been in a meeting where important decisions and action items get lost in the conversation? We built Helios to automatically transform spoken discussions into visual, interactive knowledge graphs - making it easy to see relationships, track decisions, and never miss important details.

🚀 What it does

Helios listens to your conversations through voice AI and automatically generates a live, interactive knowledge graph that:

Extracts key entities (people, systems, databases, services) in real-time
Identifies relationships between concepts as you speak
Tracks decisions and action items automatically
Organizes information hierarchically - important nodes at the top, details below

🛠️ How we built it

Frontend:

React + TypeScript + TailwindCSS for the UI
React Flow for interactive graph visualization
Vapi SDK for real-time voice transcription and AI conversation
Socket.IO client for real-time updates

Backend:

Node.js + Express server
Socket.IO for WebSocket communication
OpenRouter API with Google's Gemini Pro for intelligent graph generation
Vapi API for voice processing and transcription
Custom graph service with intelligent positioning algorithms

🔊 VAPI Integration

Real-Time Voice-to-Knowledge Graph Pipeline

1. Live Transcription & Conversation Tracking

Captured real-time speech-to-text during team discussions
Tracked full conversation history with speaker attribution
Maintained conversation context throughout entire session

2. AI Assistant Integration

Two-way dialogue: Users speak naturally, AI responds with clarifying questions
Conversational flow management via Vapi's assistant API
Dynamic conversation state handling (started, active, ended events)

3. Streaming Transcript Processing

Real-time transcript updates pushed to backend via WebSocket
Incremental graph building as conversation progresses
Periodic graph refinement every 10 seconds during active recording

4. Multi-Modal Input Support

Voice recording through Vapi Web SDK
Audio file upload for transcription (supporting async workflows)
Session-based recording management

5. Advanced Conversation Analysis

Leveraged Vapi's conversation metadata for timestamps
Speaker identification for attribution in a knowledge graph
Confidence scoring from transcription quality

6. Seamless UX Integration

Start/stop recording controls tied to Vapi lifecycle
Visual status indicators (listening, processing, finalizing)
Toast notifications synced with Vapi conversation events
Clean error handling for connection issues

Result: Users speak naturally about their ideas, and our AI instantly transforms messy conversations into structured, visual knowledge graphs - all powered by Vapi's robust voice infrastructure.

Key Technical Achievement: We went beyond basic transcription by using Vapi as the real-time data source for an intelligent graph-building system, demonstrating Vapi's capability to power complex AI workflows, not just simple voice capture.

Key Technical Features:

Smart node placement algorithm that organizes by importance (top-to-bottom layout)
Real-time position synchronization when users drag nodes
Rate limiting and throttling to handle API constraints
Collision detection to prevent overlapping nodes
WebSocket-based live updates across all connected clients

🏆 Accomplishments we're proud of

Real-time AI analysis - Processing conversations and updating graphs live with minimal latency
Intelligent layout system - Nodes automatically organize by importance with perfect spacing
Seamless voice integration - Natural conversation flows directly into visual structures
Production-ready architecture - Rate limiting, error handling, and graceful fallbacks

📚 What we learned

Robust schemas are imperative to the full-stack development process as it creates solid foundations to build off of
Bring ideas. If you can think it, the tools are there to make it
Stick to the MVP. Make something first, then make it better. Not the other way around

🔮 What's next for Helios

Persistent storage - Save and reload conversation graphs
Advanced AI models - Experiment with different LLMs for better entity extraction
Graph analytics - Insights on conversation patterns and key topics
Meeting integrations - Connect with Zoom, Teams, or Google Meet
Template library - Pre-built graph structures for common meeting types