Live demos
Six things you can try right now
Each demo states plainly what is real and what is modelled. None of them overstate what is running in your browser.
1. Voice pipeline
Real: microphone capture and RMS-based endpointing run live via
getUserMedia and Web Audio. Modelled: the transcription / LLM / synthesis stage timings are simulated, since this page has no backend. Audio never leaves your device — nothing is uploaded or recorded.Endpointing
0ms
Transcription
0ms
LLM first token
0ms
Synthesis
0ms
Total turn time: 0ms
2. Answer scoring
Real: a transparent heuristic scorer runs entirely client-side against the text you enter. Modelled: this is a rule-based approximation, not the production LLM-graded rubric.
3. Multi-agent pipeline
Modelled: this simulates agent timing and message content to illustrate the architecture. It does not call a live LLM.
Research
idleAnalysis
idleRisk
idleWriter
idle4. Architecture explorer
Real: this reflects the actual component breakdown used in the voice pipeline projects on this site.
Select a component to see what it does and why it was built that way.
5. Cost estimator
Modelled: illustrative unit costs, not a quote. Real inference pricing varies by provider and changes over time.
Monthly inference cost
$0
Cost per user: $0 · Expected turn latency: 0ms
6. Live GitHub
Real: this fetches the six most recently updated public repositories directly from the GitHub REST API at page load.
Decisions
Provider selection
| Layer | Choice | Trade-off accepted |
|---|---|---|
| Transport | WebRTC over WebSockets | More setup complexity for lower, more consistent latency |
| Hosting | Fly.io + Pipecat Cloud | Less control than bare metal, faster time to a reliable region-aware deploy |
| Endpointing | RMS threshold over learned VAD | Less adaptive to noisy rooms, easier to tune and debug per language |
Live product