Friday, May 1, 2026
About
Pulse AI is a realtime AI workspace built for engineers who need more than a response — they need to see exactly how it got there.
Every conversation runs through a production-grade streaming pipeline: tokens arrive over WebSocket, get batched at 60fps using a requestAnimationFrame scheduler, and commit to the UI at screen refresh rate regardless of model speed or response length. No dropped frames. No UI stutter. No lost tokens on reconnect.
The workspace is built around observability. A live telemetry rail tracks WebSocket latency (p50/p95), render FPS, frame commit time, dropped frames, and active streams — all updating in real time alongside the conversation itself. A built-in chaos panel lets engineers inject synthetic load: concurrent streams, network instability, reconnect storms, worst-case scenarios. A benchmark runner executes eight reproducible stress scenarios with pass/fail thresholds, so regressions get caught before users do.
Streaming resilience is handled at the transport layer. Every server message carries a monotonic sequence number. On any disconnect — network drop, tab backgrounded, mobile sleep — the client reconnects, reports its last seen sequence, and the server replays the exact gap from a Redis buffer. No token is lost. No response restarts.
Users connect their own API keys for any supported provider. Keys are encrypted at rest with AES-256-GCM, decrypted server-side per request, and never exposed to the browser after submission. Seven providers are supported: Anthropic, OpenAI, Google Gemini, xAI Grok, OpenRouter, Together AI, and Groq — switchable per conversation with per-provider model selection.
Tech Stack:
- Frontend: React 19, TypeScript, Vite, Zustand, TanStack Virtual, Tailwind v4 — deployed on Vercel
- Backend: Node.js, Fastify v5 (async) — deployed on Railway
- Database: PostgreSQL (raw pg driver, no ORM)
- Cache / Replay buffer: Redis (ioredis)
- Transport: Raw WebSocket API (browser) + ws (server), SSE fallback
- Auth: JWT + httpOnly cookies, refresh token rotation
- AI Providers: Anthropic, OpenAI, Gemini, Grok, OpenRouter, Together AI, Groq (BYOK, AES-256-GCM encrypted)
- Encryption: AES-256-GCM, per-key random IV, server-side only
- Monorepo: Turborepo + pnpm workspaces
- Testing: Vitest
- Deployment: Vercel (frontend) + Railway (backend)
