Senior AI Engineer, Voice Agent Orchestration

Clara answers the phone when a clinic can't. We're hiring the engineer who will lead how she listens, understands and decides what to say next. You'll know the frontier of voice AI well enough to tell us what to replace. We're not attached to our current design. We're attached to evidence: a new design ships when it beats the old one on our real calls. About Curo Curo builds voice agents for healthcare practices. Our first agent, Clara, handles after-hours patient calls. She captures intake, answers approved questions, escalates urgent calls and hands a clean lead to the clinic's CRM. Clara is already taking real patient calls for our first fertility-network customer—one of the largest IVF networks in the world—and we’re onboarding additional clinic groups now. We’re a small founding team building production voice AI for healthcare, and we’re looking for a founding engineer who wants meaningful ownership and direct impact on the product. Curo is based in Allentown, Pennsylvania, with a remote engineering team. The role You'll own and lead Clara's orchestration layer: everything between a caller's audio and Clara's next words. That covers turn-taking, intent classification, slot capture, dialogue control, and which models do which jobs. Today that layer is a deterministic state machine fed by a cascade of classifiers. The language model extracts facts but never routes the call. That design has earned its place on real calls, and it's yours to rethink. Two rules stay fixed, whatever you build. Weak evidence never triggers an irreversible action, like a transfer or an ended call. Safety paths fail safe. You'll build on the multi-tenant platform underneath, working with the engineer who designed it. Every new clinic brings new providers, offices, vocabulary and callers. Your job is to make adapting to them a data change, not an engineering project. This is a Senior-level role: you set the direction for this layer. What you'll do Lead the orchestration layer. Own its architecture and roadmap, including the call on what stays deterministic and what moves to models. Bring the frontier in, on evidence. Realtime and speech-to-speech models, semantic endpointing, LLM-led dialogue with guardrails: test each against real calls and adopt what wins. Build the gate. Extend our real-call replay and simulated-caller evals until they can score a redesign, not just a patch. Make understanding survive the phone line. Intent classification and slot capture for spelled names, dates of birth, phone numbers, emails and addresses, through ASR errors, corrections and readbacks, in English and Spanish. Own the feel of the conversation. Turn-taking, interruptions and latency, so callers don't talk over Clara or wait through silence. Build the loop that improves it. Mine failures from real call traces, adjudicate gold sets, distill cheaper models, and promote only what beats the incumbent on held-out real calls. Make clinic onboarding a data change. Per-clinic vocabularies, catalogs, speech-recognition keyterms and per-language wording, each with an eval gate before the first call. What you'll bring You’ve led production conversational AI. You’ve owned the runtime real people interact with—from turn-taking through understanding—and helped set its technical direction. That usually means 6+ years working in ML, NLP, conversational AI, or adjacent systems. Fluency at the frontier. You’re hands-on with current realtime and speech-to-speech models, streaming speech recognition and synthesis, and LLM tool calling. You have an informed view of what’s ready for regulated production today—and what isn’t. The phone line’s failure modes, firsthand. You know what happens to digits and spelled names under ASR, how turns get split mid-thought, how barge-in breaks conversations, and what tight latency budgets feel like in production. Evaluation as a reflex. You’ve built gold sets, regression suites, and offline replay over real conversations. You prove a new design beats the old one before switching over. Technical leadership. You’ve owned an architecture, made the call to rebuild it—or not—and brought other engineers along. Production Python and care with regulated data. You write typed, tested production code and have experience working with PHI under HIPAA or similarly sensitive data. If you match most of this, apply. We’d rather meet you than have you rule yourself out. Nice to have Healthcare voice AI: patient access, scheduling, intake or triage. Deep experience with Pipecat, LiveKit or Twilio, or contributions to open-source voice-agent frameworks. Spanish-language voice agents. Distillation, active learning or weak supervision, and LLM judges calibrated against human adjudication. Multi-tenant conversational platforms. Property-based testing, such as Hypothesis. How we work Not married to the design; married to the evidence. Every change ships with before-and-after numbers from real calls in the pull request. For a redesign, building that measurement is part of the work. Test first. The property fails before the fix lands. Real calls close tickets. A ticket closes when a real call proves it, not when the suite goes green. Failure modes are chosen. Safety paths fail safe, and no change trades a wrong answer for a crash. We build with AI coding agents every day. You'll use them to move faster, and hold their output to the same bar as your own. Our stack This is what runs today. You'll decide what the orchestration layer runs next. Voice: Pipecat (as the media layer only), Deepgram Flux, ElevenLabs, Twilio Models: Anthropic Claude (API and AWS Bedrock), bge-small embeddings, scikit-learn distilled heads, spaCy Platform and measurement: Python, AWS (Fargate, S3), Weights & Biases Weave, pytest with Hypothesis Compensation, location and benefits Base salary: $140,000–$160,000 Equity: Meaningful stock option grant, up to 2%, depending on experience and scope Location: Remote, U.S. Visa sponsorship: Not available at this time Benefits: 20 days of paid time off plus company holidays. We do not currently offer employer-sponsored health, dental, vision, or 401(k) benefits. We expect to begin expanding our benefits program in 2027. How to apply Apply with your resume and a short note, five sentences at most, answering one question: Tell us about a time you replaced part of a voice or dialogue system with something newer. How did you prove it was better before you switched over? Specific beats polished. We accept applications on an ongoing basis, so there’s no set deadline. Equal opportunity Curo is an equal opportunity employer. We welcome applicants of every background and do not discriminate based on race, color, religion, sex (including pregnancy), sexual orientation, gender identity, national origin, age, disability, genetic information, veteran status, or any other characteristic protected by applicable law. If you need a reasonable accommodation during the application or interview process, please let us know.