About the role
You will deeply understand the research workflows of our Finetuning teams and automate the high-friction parts – turning days of manual experimentation into hours. You’ll build the tools and infrastructure that enable researchers across the organization to develop, evaluate, and optimize reward signals for training our models. Yourscalable platforms will make it easy to experiment with different reward methodologies, assess their robustness, and iterate rapidly on improvements to help the rest of Anthropic train our reward models.
This is a role for someone who wants to stay close to the science while having outsized leverage. You'll partner directly with researchers on the Rewards team and across the broader Fine-Tuning organization to understand what slows them down: running human data experiments before adding to preference models, debugging reward hacks, comparing rubric methodologies across domains. Then you'll build the systems that make those workflows 10x faster. When you have bandwidth, you'll contribute directly to research projects yourself. Your work will directly impact our ability to scale reward development across domains, from crafting and evaluating rubrics to understanding the effects of human feedback data to detecting and mitigating reward hacks.
We're looking for someone who combines strong engineering fundamentals with research experience – someone who can scope ambiguous problems, ship quickly, and cares as much about the science as the systems.
Note: For this role, we conduct all interviews in Python.
Responsibilities
- Design and build infrastructure that enables researchers to rapidly iterate on reward signals, including tools for rubric development, human feedback data analysis, and reward robustness evaluation
- Develop systems for automated quality assessment of rewards, including detection of reward hacks and other pathologies
- Create tooling that allows researchers to easily compare different reward methodologies (preference models, rubrics, programmatic rewards) and understand their effects
- Build pipelines and workflows that reduce toil in reward development, from dataset preparation to evaluation to deployment
- Implement monitoring and observability systems to track reward signal quality and surface issues during training runs
- Collaborate with researchers to translate science requirements into platform capabilities
- Optimize existing systems for performance, reliability, and ease of use
- Contribute to the development of best practices and documentation for reward development workflows
You may be a good fit if you
- Have prior research experience
- Are excited to work closely with researchers and translate ambiguous requirements into well-scoped engineering projects
- Have strong Python skills
- Have experience with ML workflows and data pipelines, and building related infrastructure/tooling/platforms
- Are comfortable working across the stack, ranging from data pipelines to experiment tracking to user-facing tooling
- Can balance building robust, maintainable systems with the need to move quickly in a research environment
- Are results-oriented, with a bias towards flexibility and impact
- Pick up slack, even if it goes outside your job description
- Care about the societal impacts of your work and are motivated by Anthropic's mission to develop safe AI
Strong candidates may also have experience with
- Experience with ML research
- Building internal tooling and platforms for ML researchers
- Data quality assessment and pipeline optimization
- Experiment tracking, evaluation frameworks, or MLOps tooling
- Large-scale data processing (e.g., Spark, Hive, or similar)
- Kubernetes, distributed systems, or cloud infrastructure
- Familiarity with reinforcement learning or fine-tuning workflows
Representative projects
- Building infrastructure that allows researchers to rapidly test new rubric designs against small models before scaling up
- Developing automated systems to detect reward hacks and surface problematic behaviors during training
- Creating tooling for comparing different grading methodologies and understanding their effects on model behavior
- Building a data quality flywheel that helps researchers identify problematic transcripts and feed improvements back into the system
- Developing dashboards and monitoring systems that give researchers visibility into reward signal quality across training runs
- Streamlining dataset preparation workflows to reduce latency and operational overhead