Voice AI Engineer in Sydney

Real conversations. Real latency budget. Our agent already holds collection calls with real customers. Every pause, interruption and wrong turn is heard by someone. You own how that sounds.

A demo forgives everything. A live call forgives nothing. Our autonomous collector works live accounts today, inside Australian consumer credit rules. That means a real person answers the phone, is identified correctly, hears an accurate balance and is offered an arrangement they can afford. On a bad carrier line, while a dog barks, after a long day.

The conversation logic is the easy half. The hard half is the voice stack underneath it: how fast the agent responds, how it behaves when it is interrupted, what happens when a transfer fails, and whether we can prove a change made calls better rather than only different.

You will:

• Own the real-time pipeline: speech in, model, speech out, and every millisecond between them.
• Make turn-taking feel human: Barge-in, silence, hesitation, people talking over the agent and people who go quiet because they are upset.
• Hold the latency budget: Measure time to first audio per call, find where it goes and take it back.
• Work the telephony layer: SIP into contact-centre platforms, carrier audio quality, DTMF, transfers to a human and callbacks that actually land.
• Choose and tune the model layer: from hosted frontier models to open-weight models we serve on our own GPUs, and know why each call uses what it uses.
• Build the evaluation loop: replay recorded calls, score behaviour change before it reaches a real customer, and keep the regressions visible.
• Keep the guardrails exact where they must be: The agent has latitude to converse; identity, money and telephony are not negotiable.

The stack today is opinionated, not precious. This is what we run now. Argue for something better and bring the measurement that settles it:

• Python, asyncio, Pipecat and Pipecat Flows
• Streaming speech-to-text and neural text-to-speech
• Hosted frontier models and self-hosted open-weight models on our own GPUs
• SIP telephony into contact-centre platforms
• A mock-customer harness for replaying and grading calls

Your first 90 days:

• First 30 days: Listen to real calls and ship your first improvement to how the agent handles them. You will be in the live pipeline in week one.
• By day 60: Own a measurable part of the conversation quality, such as latency or interruption handling, with numbers that show the change.
• By day 90: Set the direction for the voice stack: what we run ourselves, what we buy and what we need to measure next.

You might be a fit if:

• Job titles vary. Voice engineer, speech engineer, backend engineer who fell into real-time audio and never left.
• You have built something that talks in real time. Pipecat, LiveKit Agents, a provider realtime API or your own pipeline; what matters is that it ran for real users.
• You are fluent in Python and comfortable in async code where the deadline is audible.
• Audio does not scare you. VAD, barge-in, jitter, packet loss and codecs are problems you have debugged rather than read about.
• You reason from evidence. You listen to the call, read the timing trace and find the cause before changing the prompt.
• You care what happens to the person on the other end. These are conversations about money with people who are often under pressure.
• Telephony scars are a bonus. So is having run models on your own hardware.

What you get:

• Production calls, not a research backlog. Your change is on live calls in days, and you hear the result. Very few voice teams get a feedback loop this short or this honest.
• The offer includes a competitive salary and meaningful equity. We will discuss the range and the trade-offs openly in the first conversation.
• Sydney is easiest, because sitting with a collections floor teaches you things a transcript will not. Remote works if your day overlaps ours by about four hours.

The process:

• Founder conversation: A direct discussion about the company, the role and what you want next.
• Practical working session: Work through a real, anonymised problem from our production call stack with the people you would work beside.
• Mutual decision: A final conversation with both founders, clear references and a transparent offer.