Voice AI engineering for production, not demos.
We design, build, and fix voice agents and real-time speech translation systems that hold up on real calls. Pipecat, LiveKit, SIP telephony, and every stage in between.
[ 01 — 09 ]
Built by the team that runs their own stack in production. Working directly with your engineers, in your stack, with clear access boundaries.
The demo is the easy part.
Most voice AI demos are built to pass a clean test call: clear audio, one speaker, a happy path. Production is different. The agent meets a real phone call: compressed audio, interruptions, a carrier that mangles the codec, a session that dies at minute nine.
Most teams spend weeks in that gap, and the vendors are no help, because every vendor's dashboard says their stage is fine.
The gap has structure. Latency lives in specific stages. Failures happen at specific seams: turn detection, barge-in, handoff, session limits. If you measure the pipeline instead of guessing at it, the fix is usually smaller than the fear.
That measuring is what we do.
Teams building voice agents who are stuck, slow, or not ready to hire full-time:
Your agent demos well but falls over on real calls, and nobody can say exactly where.
You need phone-based AI on your existing telephony, and the SIP integration is fighting you.
You are burning weeks on latency, and every vendor blames the other vendor.
Latency budgets that keep blowing out.
Human turn-taking runs on gaps of a few hundred milliseconds. Once an agent's round-trip drifts past a second, callers feel it. We trace real calls stage by stage, transport, VAD, STT, turn detection, inference, TTS, and show you in milliseconds where the budget is going. In many pipelines, most of the avoidable delay sits in one or two stages, not ten.
Telephony integration that will not behave.
SIP, SBCs, codec mismatches, carrier quirks that no documentation admits to. We have designed call translation architecture for carrier-style environments and know where the traps are before your first failed call.
Agents that pass the demo and fail in production.
Session caps that kill parallel streams. Handoffs that drop context. Degradation that nobody designed. The gap between a demo and a deployment is exactly where we work.
Where voice systems break
| Layer | What fails | What we inspect |
|---|---|---|
| Telephony | Calls fail, transfer breaks, codecs mismatch | SIP path, routing, transfer logic |
| Audio | Speech is clipped, delayed, or unstable | VAD, chunking, buffering, resampling |
| STT | Transcripts arrive late or wrong | Streaming setup, endpointing, language handling |
| Inference | Replies are slow or inconsistent | Model latency, prompts, tools, context |
| TTS | First audio is delayed or unnatural | Streaming TTS, voice path, playback |
| Runtime | Demo works, production fails | Sessions, concurrency, retries, rate limits |
| Handoff | Caller gets trapped or context is lost | Escalation rules, transfer flow, summaries |
| Monitoring | Nobody knows what broke | Traces, logs, latency by stage, alerts |
Telephony
- What fails
- Calls fail, transfer breaks, codecs mismatch
- What we inspect
- SIP path, routing, transfer logic
Audio
- What fails
- Speech is clipped, delayed, or unstable
- What we inspect
- VAD, chunking, buffering, resampling
STT
- What fails
- Transcripts arrive late or wrong
- What we inspect
- Streaming setup, endpointing, language handling
Inference
- What fails
- Replies are slow or inconsistent
- What we inspect
- Model latency, prompts, tools, context
TTS
- What fails
- First audio is delayed or unnatural
- What we inspect
- Streaming TTS, voice path, playback
Runtime
- What fails
- Demo works, production fails
- What we inspect
- Sessions, concurrency, retries, rate limits
Handoff
- What fails
- Caller gets trapped or context is lost
- What we inspect
- Escalation rules, transfer flow, summaries
Monitoring
- What fails
- Nobody knows what broke
- What we inspect
- Traces, logs, latency by stage, alerts
Trace first. Then fix. Then keep it fixed.
We do not start with opinions. We start with evidence: real calls, traces, logs, latency, SIP behaviour, vendor limits, and failure examples.
That evidence becomes the build spec.
Trace
First, we trace real calls through your pipeline and measure every stage: transport, VAD, STT, turn detection, inference, TTS. Then we map the failure modes: session limits, barge-in behaviour, handoff logic, what happens under load. You get a written report: your latency budget stage by stage, every failure ranked by impact, and a verdict on each vendor in your stack. If the fixes are simple, you will not need us again. If they are not, the audit fee comes off a Build Sprint.
Fix
We build the riskiest piece first. Telephony and integration land in week one, because the demo is the easy part and SIP is where sprints die. Mid-sprint you get a working call on your own stack and an honest go or no-go. If we find in week one that the agreed outcome is not technically reachable within the sprint scope, we stop, show the trace evidence, and do not invoice the second milestone. Week two is hardening: failure handling, handoff logic, load. You end with a system on your infrastructure and a handover doc your engineers can run without us.
Operate
Voice pipelines decay. Vendor APIs change, models get deprecated, latency drifts. We monitor your pipeline, fix what breaks, and keep it current as models and vendor APIs move. We track usage, latency, vendor changes, and error rates so cost spikes are visible early and can be handled before they become operational surprises. Every month you get the numbers: what drifted, what we fixed, what is coming.
Advisory
Your engineers build. We sit beside them on the decisions that are expensive to get wrong: which STT vendor for your accents, where the latency budget should be spent, how the handoff should work. One day of the right argument early saves a month of rebuild later.
Payment terms are agreed in the proposal after the scoping call. Sprints are typically split 50/50; audits are invoiced on booking. EUR invoicing available for EU clients.
We run this stack ourselves.
Live translation at an EU cybersecurity conference.
Leaders from across Europe, no shared floor language, and no interpreter infrastructure in the venue. We deployed SentiVue's real-time speech translation: speakers presented in their own language, attendees scanned a QR code and listened in theirs.
- 200+ participants
- QR-based access, no app download
- No interpreter booth
- Live event conditions, setup in minutes
The system on stage that day is the same pipeline architecture we audit, build, and maintain for clients.
- We have built and shipped European Portuguese speech technology, including dialect-focused voice work.
- We have designed bidirectional phone-call translation architecture for carrier-style environments, including SBC and B2BUA topologies and latency budgeting.
- We have diagnosed parallel-session failures in production Gemini Live deployments, the kind that only appears under real load.
Production stack: Pipecat, LiveKit, Deepgram, ElevenLabs, direct carrier integration.
This week
Two audit slots available
Three days. We trace your pipeline and hand you the report: latency budget by stage, failures ranked by impact.
This month
One build sprint slot open
Two weeks. A scoped build or fix shipped on your stack, with an honest go or no-go in week one.
Ready to leave
the sandbox?
You describe the problem. We tell you on the call whether we can fix it and what it costs. If we are not the right fit, we say so and point you somewhere useful.