Voice agents that actually hold a conversation.
Real-time voice AI is the hardest kind of AI product to ship well: users notice 300 milliseconds of lag, talk over the agent, and ask things you never scripted. I build voice agents that survive all three — streaming pipelines, interruption handling, and tool calling wired to your real systems — and I've shipped them to production, including a live AI interviewer that adapts to every answer.
- A production voice agent for web, in-app, or phone — speaking your business, not a generic script
- A streamed, latency-budgeted pipeline (speech-to-text → reasoning → speech) that responds like a person, not a queue
- Tool calling into your real systems: calendars, CRMs, order status, lead capture — the agent acts, not just talks
- Interruption (barge-in) handling and turn-taking that survives real humans talking over it
- Scoped abilities with graceful failure and human handoff, so the agent never improvises answers about your business
- Transcripts, analytics, and an eval suite of real conversations so quality is measured, not vibes
Conversation design & scope (week 1)
We map what the agent must do, what it must never do, and which systems it touches. You get a fixed-scope SoW with a latency budget and a price.
Pipeline build
Streaming speech pipeline, model orchestration, and tool integrations — demoed weekly with real audio, not slide decks.
Eval against real conversations
A test suite of recorded scenarios — interruptions, accents, off-script questions — scored automatically so regressions are caught before users find them.
Launch + 30 days of tuning
We ship, then tune turn-taking, latency, and edge cases against production traffic. You get transcripts and analytics from day one.
AmtHeld — an AI phone assistant answering citizen calls for German government agencies, 24/7 in 25+ languages. I worked on the voice pipeline and analytics dashboard.
Intavue's real-time voice interviewer reads your resume and asks live follow-ups — a production voice pipeline I built solo, end to end.
What it actually takes to ship a voice AI agent in production — the latency, interruption, and failure lessons first-hand.
Fixed-scope with milestones under my AI integration lane. The free 30-minute call includes an honest read on whether voice is the right interface for your use case — sometimes a text flow wins, and I will tell you.
What can a voice AI agent actually do for my business?
Anything a phone-tree or a first-line human conversation does today: qualify leads, book appointments, answer product questions, run structured interviews, or guide users through onboarding. The agent listens, speaks, and calls your real systems — checking availability, saving a lead, sending a follow-up — instead of just chatting.
How do you make it feel natural instead of like a laggy robot?
Latency budgeting, end to end. Every stage — speech-to-text, the model, tool calls, text-to-speech — gets a budget, and the pipeline streams so the agent starts responding while it's still thinking. Interruption handling (barge-in) is designed in from the start, because real people talk over agents constantly.
Which voice stack do you use?
ElevenLabs and native speech-to-speech models where they fit, with OpenAI and Anthropic models driving reasoning and tool use. The stack is chosen per project — phone, web, or in-app — and built behind a seam so components can be swapped as the voice landscape shifts, which it does every quarter.
What happens when the agent doesn't know the answer?
It fails gracefully by design: scoped abilities, explicit refusal behaviour, and a clean handoff path to a human (or a callback) with a transcript. An agent that confidently improvises answers about your business is a liability — constraint is a feature, and it's tested in the eval suite.
Have you actually shipped one of these?
Yes, repeatedly. I worked on the voice pipeline and analytics dashboard of AmtHeld, an AI phone assistant answering citizen calls for German government agencies 24/7 in 25+ languages. Intavue runs my real-time AI voice interviewer in production — it reads a candidate's resume and asks live follow-ups. And this site's own assistant can take voice conversations and book meetings. All live, none demos.
What does it cost and how long does it take?
Voice work falls under my AI integration lane: from $6k, typically 2–6 weeks depending on how many systems the agent needs to touch. It starts with a free 30-minute call and a fixed-scope SoW, and includes 30 days of post-launch support while we tune against real conversations.