Senior/Staff AI Engineer (Agents & Models)

**The most impactful thing you can do in healthcare? Delete the admin.** About a quarter of US healthcare spending, close to a trillion dollars a year, goes to administration rather than care. It's the phone that rings all day, the claim that bounces, the notes still unfinished at 9pm. Take that work away and clinicians get their time back, practices get their money back, and patients get their clinician back. **Carepatron is the operating system for healthcare practices** Thousands of clinicians run their whole practice on Carepatron: unified inbox → scheduling → EHR → billing and revenue cycle. Because we run every step, our agents can do the work end to end. A client messages on Sunday night and an agent books them in. After the session, the note and codes are drafted. The claim goes out, and if the insurer stalls, an agent chases it. It's faster, cheaper and more accurate than doing it by hand. That's what makes us different. Most healthcare AI companies have to plug into someone else's EHR before their agents can do anything, and their agents stop at its edge. We are the EHR, so our agents finish the job instead of handing it back to a person. **We're hiring an AI Engineer** to build the agents that take the admin out of healthcare: the harness they run on, the models behind them, and every form they take, from voice and browser to text and whatever comes next. **What you'll build** * The harness itself: how our agents plan, call tools, manage context and memory, ask for permission, and stay inside guardrails when they act in a clinical record. * Voice agents that answer a practice's phone and handle the call in real time: telephony, speech-to-text, text-to-speech, turn-taking and latency. * Browser agents for websites with no API, like insurer portals and the other systems clinicians are stuck logging into. * The models. When a frontier model is too slow, too expensive or not good enough at a clinical task, you'll fine-tune or train our own, and use evals to prove it's better. * The evals and monitoring that tell us whether an agent did the job right, and catch it when it didn't, before a clinician does. * You'll own each agent from prototype to production: talk to the clinicians who'll rely on it, decide what "done right" means, build it, ship it, and watch how it behaves with real clients. **You might be a fit if** * You've shipped agents or LLM features to real users, or trained, fine-tuned or distilled models that made it to production. Ideally both. Either way, you can walk us through how they failed and what you changed. * You write production code, mostly in Python, TypeScript, or C#. * You write evals as part of building, not after it, and you know when not to trust them. * You care how an agent comes across to the people it serves. **Bonus points** * Model work: distillation, fine-tuning (SFT, LoRA), preference tuning or RL, and serving small models with low latency. * Real-time voice: WebRTC/SIP, streaming speech models, handling interruptions. * Browser or computer-use agents (Playwright, computer-use models). * Healthcare, insurance or billing workflows. You don't need a healthcare background; you'll pick up the domain from us and from our clinicians.