System Architecture
A resilient, globally distributed system designed for real-time WebRTC audio/video and sub-millisecond AI inference.
Global Edge Network
WebRTC Streaming
Our WebRTC infrastructure maintains bidirectional, sub-50ms audio/video pipelines. Voice Activity Detection (VAD) models run on the edge to minimize transport payload and ensure seamless turn-taking between the candidate and the AI agent.
// Pseudo WebRTC Config
const rtcConfig = {
iceServers: [
{ urls: 'stun:global.stun.twilio.com:3478' }
],
sdpSemantics: 'unified-plan'
};State & Memory
We leverage Redis for ephemeral interview state and Supabase with pgvector for long-term memory. Context windows are dynamically managed: sliding windows for recent turns, and vector retrieval for recalling points made 20 minutes ago.
- L1 Cache: Local Browser State
- L2 Cache: Edge Redis (up to 10k context)
- L3 Storage: pgvector (infinite recall)
Security & Isolation
Code execution environments for technical interviews are strictly sandboxed via Firecracker microVMs. Each candidate session spins up an isolated, ephemeral runtime that is destroyed instantly upon interview completion, preventing cross-tenant data leakage.