Research / Data Engineer - Post-Training & RL

Researcher or engineer? This role wants both, in one person.

A profitable, seed-stage AI company in San Francisco creates long-horizon reinforcement learning environments and post-training data for leading AI labs in the US and abroad. Most data providers solve demand by adding people. This team treats post-training data as a research problem and scales it with processes that improve themselves. Demand for their data is bigger than what they can deliver today, so they're growing a small, very strong core team of researchers and engineers.

What you'll own

• The full post-training loop: infrastructure, training runs and the data that goes into them
• Long-horizon RL environments and tasks in many fields, from science to ML research itself
• Scaling data generation through smarter, self-improving tooling, not more headcount
• Judging which real-world tasks are worth building, because they matter to the labs and the economy

What you bring

• Real depth in ML research (think post-training, RL, evals or interpretability) and serious engineering skills
• You've post-trained models yourself and built the datasets for it
• A good eye for data, or the research ability to build that eye fast
• You read code fast and solve problems quickly in PyTorch and ML pipelines
• Curiosity that shows: papers, side projects, open source contributions or writing
• Roughly 3–7 years of experience

Bonus points

• Both published research and things you actually shipped
• Domain depth next to ML, such as neuroscience, physics, chemistry or systems
• Artefacts people use: fast libraries, benchmarks or well-cited papers

What's in it for you

• $250k–$300k base
• Work that goes straight into how frontier models are trained
• A small, senior team where research and engineering are one job
• Open to visa holders as well as US citizens and green card holders

Good to know

• Full-time, on-site in San Francisco, 9-9-6
• Process: first call with a founder - technical interview (a standalone problem, with an ML version in PyTorch) - work trial - offer