We didn’t build this alone.

The pioneer that made the rest of the loop possible.

ElevenLabs didn’t just ship better TTS. They moved the entire field — voice quality past the uncanny valley, voice cloning from seconds of audio instead of hours, and now Speech Engine: the first primitive that puts STT, TTS, and turn detection on one socket. Every voice product built after them inherits the floor they raised.

We tried the obvious alternative — a Whisper transcription service, a frontier LLM, and a separate TTS vendor stitched together. End-to-end latency landed near two seconds. Voices drifted. Interrupting the AI mid-sentence required custom plumbing. Speech Engine collapsed all three into a single primitive and the loop dropped to well under a second on the first try.

await elevenlabs.speechEngine.attach(SPEECH_ENGINE_ID, httpServer, '/ws', {
  onTranscript: async (transcript, signal, session) => {
    const stream = await llm.chat({ signal, ... });   // AbortSignal threaded
    await session.sendResponse(stream);                // STT → LLM → TTS
  },
});

The infra that turns “works on my laptop” into production.

Frameworks, services, and where each Dockerfile lives.

vought-os/
├── apps/
│   ├── web/                  → Vercel (marketing)
│   └── app/                  → Vercel (product)
├── services/
│   ├── echo-engine/          → Render (Node, Dockerfile)
│   └── diarization-sidecar/  → Render (Python, Dockerfile)
├── packages/
│   ├── design-system/        → tokens + global.css
│   ├── motion/               → the four signature motions
│   └── ui/                   → Container, Grid, primitives
├── docker-compose.yml        → local Postgres + Redis
└── turbo.json                → pipelines

Each service has its own Dockerfile, its own deployment target, and its own scaling story. The monorepo glue is Turborepo pipelines and a shared design-system package — nothing exotic.