Voice agents
We build complete voice agents: they answer calls in a natural voice, understand what the caller wants and sort it out on the spot (bookings, enquiries, appointments, follow-ups) or hand the call to a person. We can start from an off-the-shelf platform or build the agent piece by piece on your servers, depending on what your case needs.
The problem
Missed calls out of hours or at peak times: customers who go elsewhere.
Skilled staff answering the same questions over and over.
“Press 1” phone menus or generic assistants that sound robotic and resolve nothing.
Specialty: voice agents
Two years ago, an agent that holds a phone conversation in natural Spanish and replies in under 2 seconds was within reach of very few companies. Today it is viable for a hotel chain or a construction company, as long as these six pieces work together in real time. We have built and integrated all of them in a project with real calls.
The call comes in through your phone system or number, without changing carrier.
SIP · Asterisk · codecsCleans up the sound: echo cancellation, noise reduction and volume control.
AEC · NS · AGCTranscribes in real time and detects when the caller has finished speaking.
Streaming STTUnderstands the request, looks up your information and decides what to do.
LLM · RAG · toolsReplies in a natural voice, and stops talking if the caller interrupts.
TTS · interruptionsEvery turn is traced: latency, quality and errors, so it improves with real calls.
Tracing · metricsOur approach
Some platforms can set up an agent in an afternoon, and some cases need more than that. We know both options from the inside, so we choose the one that fits your volume, your budget and your data, and change it if your case evolves.
When speed and cost come first, we start from existing platforms and do what they don't: connect them to your systems, your data and your rules.
For validating quickly or moderate volumes.
ElevenLabs · OpenAI · Deepgram · n8n
The best models available with a custom layer of orchestration, tools, quality checks and traceability. The critical parts stay in your hands.
When the AI has to work on your real processes.
Claude · GPT · Bedrock · Langfuse
Every piece built and tuned, with open language models on your own GPU. Full control over privacy, latency and cost.
For sensitive data, high volumes or requirements nobody else covers.
vLLM · Mistral · Whisper · Asterisk
Many projects start by integrating and only move to custom development where it adds value. No lock-in to any provider.
How it works
Turns the caller's speech into text in real time, in Spanish and other languages depending on the case.
A language model works out the request and looks up the information it needs.
Records the booking or enquiry, or transfers the call to a person along with the context.
Answers in a natural voice in under 2 seconds, and can be interrupted just like a person.
What's included
What you get
Technology
Where it fits
Example
Bookings and common questions outside reception hours.
Example
Appointment management, with the data kept on your servers.
Example
A first line that handles the routine and passes on the rest.
Related project
< 2 s
response time, end to end
Business call centre
A complete Spanish-language voice agent, connected to the phone system and tested on real calls, with no data leaving the client's network.
View the projectFrequently asked questions
No. It uses natural voices and is tuned for fluent conversation: it replies in under 2 seconds and can be interrupted.
If that's the most viable option, yes. But we also build the complete agent (telephony, audio, speech recognition and synthesis, an open language model on your GPU) when you need more privacy, higher volume or a level of control those platforms don't give you.
Our production experience is in Spanish. Other languages, such as English, are configured and validated during the pilot with test calls.
It can run in the cloud or on your own servers. In the on-premises version, neither the audio nor the data leaves your network.
A 30-minute conversation is enough to know whether we can help and how. No commitment.