Engineering Intern at San Francisco Data

San Francisco Data turns production traces into training environments for frontier models. We've reached $100M+ in annualized revenue run rate and raised a $30M seed round led by Khosla Ventures.

Our founders previously grew Mercor's code team from $0 to $500M in annualized revenue run rate, leading its largest code engagements and synthetics work. Our team brings experience from Google DeepMind, Etched, Nvidia, Stanford SAIL, and Berkeley RISE, including TerminalBench co-authorship.

We're a small team of engineers and researchers working in-person in FiDi, 5 days a week.

What You'll Do

• Staffed on the same client engagements as our full-time engineers and will collaborate directly with production-trace providers and researchers at frontier labs
• Develop new training environments and task formats for frontier models using production agent traces
• Analyze agent behavior and failure cases, and run experiments on new benchmarks and task formats
• Develop tools to generate, test, and evaluate training tasks, helping improve the quality and reliability of our environments
• You'll get the same level of ownership and freedom as an FTE, with matching expectations for high rate of learning and exceptional output . For the right person, this role may lead to an offer to join our technical team full-time.

What We're Looking For

• Significant contributions towards reinforcement learning and evals research (or exceptional contributions in other fields) / experience building RL environments. Work on public benchmarks is a plus.
• Prior experience building open-source projects, AI agents, or synthetic data pipelines, with a strong understanding of the challenges involved in training and evaluating frontier models.
• Proven ability to independently tackle complex technical problems, build robust systems, and design rigorous experiments to evaluate performance and drive improvements.
• Curiosity about how AI systems learn and fail, and an interest in building more effective ways to train and evaluate them.

We care more about what you've built, how you think, and your willingness to learn than specific credentials.

Apply

Apply through this page with your resume and links to benchmark work, published papers, codebases, or other evidence of exceptional talent. Tell us what you personally contributed.