Research Engineer, Post-Training
Remote • San Francisco • FullTime
Posted 2mo ago
Job Location
San Francisco
Tech Stack
Remote Work Policy
Fully remote
Employment Type
FullTime
Categories
AI Research Engineer
About the job
We are an applied AI lab building end-to-end software agents, notably the creators of Devin, the first AI software engineer. Our team is composed of highly talented individuals with backgrounds from leading AI companies and competitive programming. We are tackling significant challenges in AI, aiming to build systems that can reason and perform real-world tasks. This role focuses on post-training, which is crucial for transforming raw model capabilities into useful, safe, and effective real-world systems. You will be instrumental in shaping how our agents learn through iterating on training recipes, evaluations, and alignment methods, directly impacting the functionality of Devin and future systems. This position requires a blend of deep research and hands-on engineering, with no distinction between the two.
Responsibilities
- Iterate on datasets, training stages, and hyperparameters to define model behavior.
- Develop and refine evaluation metrics to accurately assess model performance.
- Investigate and understand the root causes of unexpected training outcomes.
- Apply and advance alignment techniques such as RLHF, RLAIF, and constitutional approaches.
- Measure performance scaling with data and compute, and develop new methodologies.
- Contribute to research and product development in a fast-paced environment.
Requirements
- Demonstrated experience in advancing ML systems through post-training, alignment, or related methods (e.g., RLHF, RLAIF, preference modeling, reward learning).
- Strong foundational knowledge in probability, statistics, and ML theory.
- Ability to analyze experimental data to distinguish real effects from noise and bugs.
- Evidence of original contributions, such as publications at top venues or significant open-source impact.
- Experience with large-scale distributed training and associated debugging.
- Systems-level thinking, understanding the interplay of training pipelines, data, and evaluation.
- Comfort working in ambiguous and rapidly evolving research environments.
- Demonstrated capability is valued over formal credentials; a PhD is one signal among many.
Benefits
- Compute is not a constraint: large allocations with training jobs routinely running across thousands of GPUs.
- Minimal process overhead, rewarding speed, autonomy, and technical depth.