The opportunity
As a Software Engineer on the RL Data team, you’ll design and build the tools that researchers and external contributors use to create, review, submit, and monitor the environments and tasks behind Cursor’s reinforcement-learning runs. This is a full-stack product-engineering role embedded in a research team.
What you'll do
Create fast, trustworthy workflows for vendors and research team to interact: effectively with each other — vendor task creation and iteration, vendor submissions, and task acceptance into training.
Build review tools for inspecting and comparing rollouts, transcripts, grader: outputs, and other signals of task quality.
Develop environment-health, failure-search, versioning, and catalog: experiences that make training data easy to understand, manage, and extend.
Establish a shared component kit, then use it to build self-serve interfaces: for creating and improving tasks with quality checks inline.
You’ve shipped full-stack products and owned systems from user interface: through storage or services, using technologies such as TypeScript and React alongside Node, Python, or Go.
You’ve built dense, data-facing tools such as transcript viewers, diffing: systems, review queues, observability products, or operational dashboards—and you have strong opinions about how structured data should be rendered.
What they're looking for
- You’ve built or maintained a design system or component library and can: establish durable product and engineering conventions for a fast-moving team.
- You’ve designed review, QA, moderation, fraud, or acceptance workflows where: users had an incentive to get past the checks, and you know how to keep those systems honest.
- You care about data quality, and are willing to inspect raw data. Experience: with evaluations, graders, reinforcement learning, or data-quality systems is helpful but not required.
- You move quickly under ambiguity, collaborate closely with researchers and: domain experts, and take open-ended problems from rough need to reliable product.
