The main responsability for this role would be to assess an existing architecture to give recommendations and implement improvements.
- Has put a voice agent on a real phone number in production using LiveKit Agents or Pipecat, with SIP trunking they configured themselves. Ask which carrier and what broke — a real one has a story about one-way audio or codec negotiation.
- Telephony audio literacy: G.711/μ-law, 8 kHz narrowband, Opus transcoding, resampling. This is the fastest filter in the whole screen.
- Endpointing and turn-taking as a subject they have opinions about — VAD versus semantic/waveform turn detection, and the false-cutoff-versus-latency tradeoff as a tuning surface rather than a setting.
- Reads framework source, not just docs. Our fix was found in audio_recognition.py, not in a config page. Someone who has filed or read GitHub issues in these repos is worth more than someone with a longer CV.
- IVR instincts — DTMF, digit grammars, slot filling, retries, confirmation read-back. The whole industry answer to digit capture came from that world, and it's a generation older than the voice-AI crowd.
- Spanish-language voice experience, ideally Caribbean or LatAm. Nice to have, and the pool shrinks fast if you make it mandatory — you already have that knowledge in-house.