SpaceX was founded under the belief that a future where humanity is out exploring the stars is fundamentally more exciting than one where we are not. Today, SpaceX is actively developing the technologies to make this possible, with the ultimate goal of making life multiplanetary and expanding consciousness to the stars.
SpaceXSI is building Super Intelligence to help humanity understand the universe. Our teams are building the full Super Intelligence stack: from our Colossus supercomputers, to Grok models, to our suite of applications, including Grok Bot, Grok Build, Grok Imagine, Grok Voice, and X. We're creating a powerful flywheel where more compute trains better models, better models power more useful products, and what we learn from people using those products shapes what we build next.
The data behind a model shapes it more than almost anything else. On the RL Data team at SpaceXAI, you decide what that data is for our agents: which problems they practice, what world they practice in, and how they're judged. You'll take a weakness, like an agent that gives up on hard bugs or misuses a calendar API, and turn it into training data that fixes it, then show the next model is measurably better.
You form hypotheses about why data is helping or misleading the model, build the systems to test them, run ablations, and read the rollouts to find the bug, shortcut or mismatch behind a result. Every flaw you catch, or miss, shows up in what the next Grok models learn.
What you’ll work onTurning a capability gap into thousands of verifiable tasks — for example, planting subtle bugs deep in a real project's dependencies for the agent to root-cause.
Building environments that behave like production: real codebases with their real dependencies, and simulations of apps like email, calendar, and docs that match the real ones down to search, errors, and pagination.
Designing graders for open-ended work that separate good solutions from bad without penalizing valid alternatives or rewarding shortcuts.
Building automated checks that catch reward hacking, leaked answers, broken environments, and paths the real product wouldn't allow, before data reaches training.
Scaling from a thousand hand-checked tasks to millions, with filters and judges you can trust in place of a person reading each one
You have strong software engineering fundamentals, and you've built non-trivial systems in Python: data pipelines, test or eval infrastructure, sandboxes, or backend services.
You have taste for what makes an environment trustworthy: faithful enough that skills transfer, strict enough that shortcuts fail, and simple enough to debug when a result looks wrong.
You are careful and systematic — understanding both aggregate metrics and close reading of raw outputs.
You enjoy turning messy, open-ended problems into repeatable, high-quality data production systems.
RL or ML experience is a plus but not required. We welcome strong engineers who want to do research on what makes models better.
If there appears to be a fit, we'll reach out to schedule 2-3 short technicals. After, we'll schedule an onsite in our office, where you'll work on a small project, discuss ideas, and meet the team.
Skills Required
- Strong software engineering fundamentals
- Experience building non-trivial systems in Python
- Experience with data pipelines, test or evaluation infrastructure, sandboxes, or backend services
- Ability to design trustworthy environments and systematic data production systems
- Careful, systematic analysis of aggregate metrics and raw outputs
- Reinforcement learning or machine learning experience
What We Do
We'd like to automate coding. To advance that mission, we're building Cursor. Our work includes training the world’s most widely used coding models, creating infrastructure that supports billions of requests per day, and building better ways for humans and AIs to work together.
