New York City · Full-time

Build the environments that teach us what healthcare AI can do—and where it needs to improve.

At Nolla Health, we’re building AI-powered healthcare. To improve the systems involved in delivering care, we need realistic ways to test them: cases that unfold over time, tools an agent can use, and clear criteria for judging what happens.

We’re looking for a research engineer to build evaluation harnesses, simulated environments, and data pipelines that turn clinical knowledge and appropriately governed data into useful evaluation and training assets.

You’ll work with engineers and clinicians to define tasks, build the systems that run them, and make the results trustworthy. The work will support our own agents and models, with the opportunity to develop benchmarks and environments for external AI teams.

What you'll do

• Build harnesses that run agents through reproducible tasks, capture their actions, and make failures easy to investigate.
• Develop multi-step and multi-visit environments with patient state, clinical tools, and realistic consequences for an agent’s decisions.
• Work with clinicians to turn clinical workflows into well-defined tasks, rubrics, and success criteria.
• Build data pipelines for preparing, validating, versioning, and tracing evaluation and training datasets.
• Develop scoring and reward mechanisms, combining programmatic checks, model-based grading, and clinician review where appropriate.
• Test graders and environments for ambiguity, shortcuts, data leakage, and reward hacking.
• Run experiments, analyze trajectories, and identify whether changes improve performance across meaningful slices of cases.
• Work with our applied AI engineer to turn evaluation findings into product improvements.
• Build with privacy, access controls, and permitted data use in mind.

What you bring

• Strong programming skills and experience building reliable software for experiments, data processing, or agent systems.
• Hands-on experience building an agent evaluation harness, an RL environment, or closely related infrastructure.
• The ability to translate an open-ended problem into a runnable task with measurable outcomes.
• Careful judgment about data quality, reproducibility, grading reliability, and what a benchmark result actually supports.