Applied AI
Voice AI & Conversational Interfaces
Real-time voice agents for phone and in-app, with low latency, barge-in, live tool calls, and a clean hand-off to a human.
Speech models are finally fast enough to hold a natural conversation, which makes voice a real interface again, for support lines, scheduling, qualification, and hands-free field work. The engineering bar is latency: under roughly 800ms round trip a conversation feels human, and above it feels broken. We build the whole path: telephony or WebRTC, streaming speech recognition, an agent that can call your systems mid-sentence, and speech synthesis, all with barge-in, interruption handling, and a warm transfer to a person whenever the caller asks.
What's included
- Sub-second response
- Telephony & in-app
- Live tool calls mid-call
- Warm human hand-off
What you get out of it
- Calls answered instantly, at any hour, in the caller's language
- Structured notes and CRM updates written during the call, not after it
- Full transcripts and recordings for quality review and compliance
Latency is the whole user experience
A voice agent lives or dies on the pause before it answers. We stream everything. Partial transcripts feed the model before the caller stops speaking, the first synthesised syllables play before the sentence is complete, and tool calls run concurrently with speech. Barge-in cancels output the instant the caller interrupts, so nobody has to wait politely for a machine to finish.
call: +55 · connected · lang=pt-BR (auto)asr: streaming partials @ 120msintent: reschedule_appointment (0.97)tool: calendar.find_slots(next_7d) -> 4 optionstts: first audio out @ 640msbarge-in: caller interrupted @ 1.2s -> output cancelledhandoff: warm transfer to agent + live summary
What we build
Phone agents
Inbound and outbound voice agents on real phone numbers, integrated with your telephony provider, IVR, and call routing, including after-hours coverage and overflow.
In-app & web voice
WebRTC voice interfaces inside your product or site, with visible transcripts and a smooth fall back to typing whenever the caller prefers it.
Live tool use
Agents that look up an order, check availability, or write a CRM note while still on the call, so the outcome is done by the time the caller hangs up.
Multilingual conversation
Language detected and switched mid-conversation, which matters for Brazilian teams serving both Portuguese and English speakers on the same line.
Transcripts & QA
Every call transcribed, summarised, scored against your quality rubric, and searchable, turning a call archive into something you can actually learn from.
Technologies we work with
- OpenAI Realtime
- Deepgram
- ElevenLabs
- LiveKit
- Twilio
- WebRTC
- Python
- TypeScript
Frequently asked questions
Yes, we disclose it, both because several jurisdictions require it and because it sets expectations that make the call go better. Callers can ask for a human at any point and get one.
Typically 600–900ms from end of speech to first audio, depending on the region and how much tool work a turn requires. We instrument it per turn and treat regressions as bugs.
In most cases yes. We integrate over SIP or through your provider's API, so you keep your numbers, routing, and recording setup.
Consent prompts, configurable retention, and redaction of sensitive fields in transcripts are part of the build, aligned to LGPD and GDPR expectations.
Next step
Need voice ai & conversational interfaces?
Tell us what you're trying to build. We'll give you an honest read on the approach, the effort, and what it costs to run.