Your voice agent works in the demo and dies in production.
Real-time voice AI engineering: Pipecat and WebRTC voice agents built and measured for the network conditions a demo never shows you.
Three signs latency is the actual problem
- Turn latency is fine on your office wifi and unusable on a phone network three time zones away.
- The agent works until two people call at once, then the audio pipeline falls over.
- Nobody on the team can say the p50 or p95 turn latency number without guessing.
How a voice agent gets to production
Measure before building
Every stage of the pipeline — endpointing, transcription, reasoning, synthesis — gets its own latency number before anything is optimised, so effort goes where it actually matters.
Build on Pipecat or an equivalent real-time framework
Not a custom WebSocket loop reinvented per project — a framework built for interruption handling, voice activity detection and streaming, so the hard parts aren't rebuilt from scratch.
Design endpointing per language
Silence thresholds and turn-taking logic tuned per language, not a single number that happens to work for English.
Deploy where real-time media actually works
Hosting chosen for WebRTC and media relay behaviour, not a general-purpose web host that happens to run the app.
Load test concurrent calls before launch
The failure mode that never shows up in a solo demo — two calls at once — gets tested before a client's users find it first.
Muhawir: sub-800ms, bilingual, in production
A bilingual voice interview practice agent running mock interviews end to end — capture, understand, respond, evaluate, report — with scoring deliberately kept out of the conversation loop so it never adds latency to the spoken turn. Deployed on Fly.io with Pipecat Cloud handling the real-time media.
Concrete deliverables
A measured latency budget
Per-stage p50 and p95 numbers, not a single end-to-end guess.
A deployed, working endpoint
Not a local demo — a hosted voice agent your users can actually call.
Interruption and turn-taking handling
The agent stops talking when the user starts, and resumes cleanly.
Load-tested for concurrency
Verified under multiple simultaneous calls before launch, not after a complaint.
The two questions every team asks
"Won't latency always be a problem with real users?"
It's a bounded problem, not an open one. Endpointing, transcription and synthesis each have a measurable floor — the work is choosing the right provider and pipeline shape for each stage, then verifying the combined number under real network conditions rather than assuming it.
"Where does this actually run?"
Hosting is chosen specifically for real-time media handling — general-purpose web hosts frequently add latency or drop connections under WebRTC load. Fly.io and Pipecat Cloud are the combination used on Muhawir; the right choice depends on the project's traffic pattern.
A voice agent build starts at $6,000
The Build tier covers a full voice or agent system end to end, deployed, across 4–8 weeks. Full tiers and payment terms are on the pricing page.