Research Engineer, Knowledge Foundations

About the role

The Knowledge Work team builds the training environments and evaluations that make Claude effective at real-world professional workflows — searching, analyzing, and creating across the tools and documents knowledge workers use every day. As that work scales, the systems behind it need to be as rigorous as the research itself.

As a Research Engineer on Knowledge, you'll design and run experiments that improve how Claude searches, retrieves, and reasons over information at scale. The work spans environment design, data curation, RL training, evaluation, and the infrastructure that supports it all. You'll move fluidly between these depending on what's blocking progress. You'll partner closely with researchers and other RL teams to ship capabilities that show up directly in Claude's behavior.

As our training and evaluations continue to scale, we see a strong synergy between the capabilities our models learn, the tools we build for them to use, and the tools we build for ourselves to understand it all. We own the science behind superhuman epistemics and we ensure the quality of the stack that drives it. We understand that real ownership and impact comes as much through hardening and iterating on environments as it does creating new ones. 

Responsibilities

• Design, build, and iterate on training environments and data pipelines that improve Claude's ability to reason over knowledge-intensive tasks
• Run experiments end-to-end: form a hypothesis, build the infrastructure, train models, analyze results, and decide what to try next
• Develop evaluations that meaningfully capture progress on search, retrieval, and reasoning quality
• Identify failure modes in current model behavior and translate them into concrete training signals
• Collaborate closely with researchers across RL Data, post-training, and product teams to align on priorities and ship improvements
• Contribute to shared infrastructure and tooling that compounds the team's velocity over time
• Own a clean, canonical set of evaluation tools and processes for Knowledge Work capabilities, including the process used for model releases
• Build and automate observability, dashboards, and operational tooling for our training environments and evaluation systems, with an emphasis on high signal-to-noise: a small set of trusted metrics and alerts rather than sprawling instrumentation

You may be a good fit if you

• Are a highly experienced Python engineer who ships reliable, well-instrumented code that teammates trust in production
• Experience designing, running, and analyzing ML experiments
• Ability to work across the stack — from data pipelines to model training to evaluation
• Have 5+ years of experience operating ML or distributed systems at scale
• Comfort working with ambiguity and choosing the most impactful problem to tackle next
• Clear written and verbal communication, especially when collaborating across time zones
• Find genuine satisfaction and impact in making existing critical systems dependable

Preferred qualifications

• Hands-on experience training, fine-tuning, or doing RL on large language models
• Experience building evaluations for LLMs, particularly in open-ended or knowledge-intensive domains
• Prior work in a research-heavy environment such as a frontier AI lab, quant research firm, or domain-focused AI startup
• Published research on LLMs, RL, retrieval, or related areas
• Experience with distributed training systems
• Are comfortable being the long-term, context-rich owner of a system and its operational health

Representative projects

• Building a training environment that teaches Claude to plan and execute multi-step research tasks against real document corpora
• Designing an evaluation suite that distinguishes genuine reasoning over evidence from plausible-sounding pattern matching
• Scaling long-running evals and fickle training environments that use many different tools
• Curating and validating a high-quality dataset of expert research workflows for use in post-training
• Diagnosing why Claude fails on a class of long-horizon retrieval tasks and proposing a training intervention, tool, or infrastructure change to fix it