Research Engineer, Post-Training

Job Details

• Experience: 3+ years
• Salary: Not specified
• Equity: Competitive equity
• Visa sponsorship: Not available
• Employment type: Full-time
• Work type: In-person

Benefits & Perks

• Health insurance and standard benefits
• Lunch covered in the office
• Support for publishing research papers, technical writing, and open-source contributions
• As an early employee, help shape the technical direction, benefits, and team activities

About This Role

As one of the first research engineers, you will own the data, evaluations, and post-training methods that help small specialized models outperform open models on real product tasks, such as software testing and domain-specific web research. You will work directly with the founders, systems engineers, and customer teams, taking each experiment from diagnosing where an agent fails to determining what evidence demonstrates improvement, and then delivering a trained and distilled model running in a customer product. The work spans the full stack, from data and benchmarks through training and inference. Successful approaches become reusable methods and tools across customer engagements.

Responsibilities

• Design and run post-training recipes that combine supervised fine-tuning, reinforcement learning, and distillation, and measure their effects on capability, generalization, and efficiency.
• Build synthetic data mixtures from production tasks and feedback, including teacher selection, trajectory sampling, and curation.
• Develop agent harnesses, environments, reward models, and verifiers, and set up benchmarks that capture real product improvements.
• Partner with systems engineers on experiment and inference efficiency, connecting algorithmic choices to rollout cost, memory use, and training stability.
• Carry promising research into customer deployments and share findings through technical writing, papers, and open-source contributions.

Requirements

• Hands-on experience training or post-training models in industry or a university research lab, with practical depth in at least one area such as RL with verifiable rewards, policy optimization (such as GRPO), on-policy distillation, fine-tuning, reward modeling, or synthetic data.
• Working knowledge of how PyTorch or JAX execute training, with the ability to read research code, diagnose training problems, and check that an implementation matches the intended method.
• Strong experimental judgment, including choosing appropriate baselines, running controlled comparisons, and paying attention to generalization, data leakage, and misleading rewards or metrics.
• Ability to explain precisely what you personally built or investigated on past projects and the evidence behind the results.