Senior Software Engineer

Parakeet Health is hiring a Senior Software Engineer to build and expand its AI agent platform for medical practices. The role focuses on shipping customer-facing healthcare workflows, improving agent conversations and reliability, and building infrastructure that enables agents to work safely with sensitive patient data.

Responsibilities:

• In your first two weeks, you’ll launch a new end-to-end agent use case for a live customer
• The gray-area decisions around patient experience are the ones LLMs consistently fumble: what the agent says when something goes wrong, how it recovers mid-conversation, when it hands off to a human. Someone with actual judgment has to own those calls, and that someone is you
• Your job, more broadly, is to harness LLMs to give yourself as much leverage as possible across everything you do. Use them to uncover product gaps, pressure-test your own ideas, and compress the distance between a thought and a shipped thing
• You’ll also be closer to customers than most engineers at your level, because the best product decisions here come from engineers who heard the problem firsthand
• Getting to this point required a lot of upfront investment, including building out our harness with rich context graphs and MCPs, and we’re starting to enjoy the fruits of that. But we've only scratched the surface, and we know it
• Getting us there is part of the job
• Building on and expanding that platform is the work
• Voice and conversation design at scale. Our agents speak to tens of thousands of patients daily. The gap between “technically correct” and “actually human” is enormous, and it’s not a problem you can prompt your way out of. One of our engineers spent three or four hours generating iterations of filler phrases (“mmm, let me look that up for you”) because cadence and intonation matter to a patient calling about their health. An LLM can generate the options. Someone with ears has to choose
• Knowing when the agent should stop. Healthcare conversations move through predictable stages—intake, verification, scheduling, follow-up—and what’s appropriate to say at each stage isn’t always obvious from the outside. A patient’s diagnosis might be relevant at one point in the conversation and completely out of bounds thirty seconds earlier. A model optimizing for correctness will flatten all of that. Building systems that know not just •what• to say but •when•—and that route edge cases to a human without grinding everything to a halt—is some of the most consequential work we’re doing. Get it wrong and it’s not just a bad experience. Depending on what gets said to whom, it can be a compliance violation
• Building the infrastructure that makes all of this possible. Getting an agent to fix a bug autonomously sounds simple. Getting it to do that reliably, on a production codebase that handles real patient data, is a different problem. It requires evaluation systems that can tell the difference between a good fix and a confident wrong answer, context that’s rich enough for the agent to understand what it’s touching, and guardrails that fail safely when it doesn’t. We’ve built a lot of this. There’s more to build. And the honest truth is that the agents are only as trustworthy as the humans who designed the systems around them
• You're building out a new capability on the agent platform: a way to handle prescription refill requests that a practice flagged as fully manual and error-prone
• An alert surfaces from the Datadog agent: something’s off in how insurance information is being collected in a specific scenario. The symptom points to one place; the root cause turns out to be somewhere completely different. The agent had context across the entire codebase, meeting transcripts, and Notion pages and traced it faster than any engineer who knew the stack. You verify the fix and ship it
• One of the QA folks who audits patient calls posts in Slack that she needs a “Mark as Reviewed” button. It takes a few seconds per call, but she reviews thousands of them. She tags the Slack bot hooked to our codebase. Five minutes later the button exists. You give it a once-over and it’s live
• You sync with CSMs and Eric on what they heard from customers. One conversation surfaces a new product opportunity. You sketch what a prototype might look like