LiveKit Voice AI Engineer — Python Phone Agents (SIP/Twilio, IVR Navigation, Prompt Engineering)

⚠️ READ FIRST — applications without a Loom video will NOT be considered. A 60-second Loom is required (details at the bottom). No Loom = automatic pass, no exceptions. Please don't send a text-only proposal. If you have never shipped a LiveKit Agents (or equivalent voice-agent) project to production, please don't apply — this is a hardening job, not a learning job. We run a healthcare SaaS for audiology clinics — live on AWS. Our AI voice agents phone insurance companies, navigate their IVR phone trees, talk to live reps, and verify patient benefits. The agents work today in production on a hosted platform; we've built a LiveKit Agents port (Python) that already completes real end-to-end calls — including a full warm transfer to a human biller on a live insurance call. Your job is to take that working scaffold and make it bulletproof: harden, test, tune, and extend it across many more insurance payers. Stack you'll live in: LiveKit Agents ~1.x (Python 3.10+) — the agent worker, tools (DTMF, SIP-REFER warm transfer, reason-gated hangup), Silero VAD LiveKit Inference / BYOK plugins — Deepgram STT, Cartesia Sonic-3 TTS, Anthropic Claude as the LLM Twilio SIP trunking — outbound trunks into LiveKit, caller ID, DTMF, warm transfers; you'll own getting our telephony fully sorted Endpointing / turn detection tuning — long hold music, payer IVRs that talk over you, "are you a bot?" challenges, speak-first timing on transfers Prompt engineering — each insurance payer's phone tree is different; we have per-payer playbooks and you'll be writing/refining agent prompts for payers we've never called before What the work actually is: Run structured live test calls against real payer IVRs, log outcomes against defined pass bars (we have a test-log discipline — you'll keep it up) Fix reliability issues with deterministic code wherever possible (evidence-gated tool calls, retry rules, anti-mishear handling) rather than prompt-only patches Tune STT/TTS/VAD/endpointing settings for the ugly realities of insurance phone lines Build out prompts + playbooks for new payers as new clinics onboard (each clinic brings payers we haven't mapped yet). Eventually help wire the agent behind our app's engine-agnostic adapter so we can flip an env var between engines. Why this is a clean job, not build-from-scratch: The agent, dialer, SIP trunk setup scripts, prompts, and test harness already exist and have completed successful live calls. Flow logic is proven: IVR navigation, member-ID/DTMF entry, benefit capture, live-rep handoff, warm transfer — all demonstrated on real calls. Day one you get: the codebase, a written hardening plan, the full test log of every call made so far (with the fixes each failure produced), and per-payer playbooks. This isn't one-and-done. After hardening ships we need ongoing work — new payers, monitoring, tuning, new capabilities. If you're solid and easy to work with, this becomes a steady long-term relationship. You're a fit if you've built and shipped voice agents on LiveKit Agents (or Pipecat, NOT just Vapi/Retell) with deep SIP + Python experience and can ramp on LiveKit fast). Tell us what you've shipped and where it ran. Compliance: HIPAA app. All test calls use synthetic/test patient data — no real patient data ever. Signed confidentiality / BAA before access. How to apply — Loom required, no calls: Record a 60-second Loom (mandatory — no Loom, no consideration): The voice agents you've shipped — platform (LiveKit? Pipecat? other?), what they did, whether they made real phone calls (SIP/Twilio) in production. Your approach to making a phone agent reliable on hostile IVRs — endpointing, turn detection, DTMF vs speech, transfer handling. When you can start + your honest weekly availability. Start your written proposal with the word TRUNK so we know you read this. We pick from the Loom videos