Sovrano AI, SL is a Barcelona company selling verified European expert capacity to frontier AI labs, robotics companies and enterprise AI teams. A network of 400,000 or more professionals, 100 or more partner universities across Europe, EFMD certification, 23 languages, a funding round led by b2venture, and a proprietary benchmark in EuroExec.
The job
You build the interview agent: a real time voice system that talks to candidates, listens, and holds a conversation that has to feel right on the first attempt. Speech to text on the way in, an LLM in the middle, text to speech on the way out, and a latency budget measured in hundreds of milliseconds. Someone interrupts halfway through a sentence and the agent handles it the way a person would. A model stalls, and the failure sounds like a pause rather than a crash.
Real time voice is unforgiving in a way that ordinary product work is not. There is no loading spinner. It either feels like a conversation or it does not, and the difference is measured in milliseconds you have to go and find.
The interface is as much of the product as the model behind it. You lead on the UI, and you care how it feels rather than only whether it works.
The stack
The agent is Python on LiveKit, with streaming speech to text and streaming text to speech either side of an LLM. The interface and the wider platform are Node.js, TypeScript and React, on PostgreSQL and Google Cloud. You work across both halves.
You own
The interview agent end to end: the real time voice pipeline, the model integration, and the prompt and evaluation loop behind it
Latency, interruption, endpointing, turn taking, and what happens when a model or a network stalls mid sentence
The candidate facing interface, from interaction design through to shipped frontend
Model quality, measured rather than asserted. Eval datasets, regression testing on prompt changes, and knowing whether a change made the conversation better or only different
Integrations that bring the agent, its output and its evidence into the rest of the platform
The experience we require
Five or more years building production software, full stack, with strong frontend depth in TypeScript and React
Real time audio shipped to production. Streaming speech to text and text to speech, WebRTC or an equivalent transport, and the work that comes with them: voice activity detection, endpointing, turn taking and barge in
LLM APIs in production. Streaming, tool calling, structured output, first token latency, cost, and what the product does when the model answers wrongly in front of a user
You measure latency rather than estimate it. A p50 and a p95 you can quote per stage of the pipeline, and tracing good enough to say which hop is the expensive one
Genuine UI and UX judgement. You take a rough idea to a designed, shipped interface and can explain the choices
AI native. You already work with AI tools daily and build what you need without being shown how
English at B2 or above
Also worth having
LiveKit, or another real time agent framework run in production
A view on the providers at each stage and why: Deepgram, AssemblyAI or Whisper for speech to text, ElevenLabs or Cartesia for speech out
Speech to speech models, OpenAI Realtime or Gemini Live, and a position on when collapsing the pipeline is worth the control you give up
Async Python and FastAPI
Evaluation tooling: Langfuse, LangSmith, or something you built yourself
Voice treated as personal data: redaction, retention and access control
