What we're looking for
We're looking for an engineer to own our voice AI end to end: the calling pipeline, the quality of every conversation that runs through it, and what we build next. Engineers take a company goal, find where the leverage is, turn it into a roadmap, ship, and measure whether it moved the number.
Our problem space runs on the phone. Getting a customer set up takes a real conversation, and much of the work we do on their behalf can only happen by calling institutions — which means automated phone menus, long holds, transfers, and getting a specific outcome out of someone who didn't expect the call. Many of our customers are older and expect to be spoken with, not processed. Voice AI has to clear that bar, and clearing it is the product.
You'll own the full real-time stack and the systems that make it measurably better with every release: evaluation for calls, replayable regression suites, and the tooling that tells us why a call went wrong before a customer does. You'll work directly with the founders and with the team that listens to these calls every day.
This role is specifically for someone with real voice experience. If you've shipped a production real-time voice agent, we want to talk.
Qualifications
- You've built and shipped real-time voice AI to production — not a demo, but a system that took real calls with real people.
- Hands-on depth across the voice stack: streaming speech-to-text, text-to-speech, turn detection and endpointing, barge-in, and latency optimization under a real-time budget.
- Experience running telephony in production (SIP/PSTN, Twilio or similar) and the failure modes that come with it.
- You've built evaluation for voice or conversational systems.
- You've debugged AI systems in production and can trace a bad outcome through audio, transcript, and model behavior to the actual cause.
- You have product judgment, and can show it: a time you challenged a spec, cut scope to ship what mattered, changed course after watching real users, or built something nobody asked for that turned out to matter.