You'll work across the lab on agent capability evals, benchmark design, LLM-as-judge systems, failure analysis, and the infrastructure that ties it together. This is a high-growth, high-ownership role on a small team, and you'll ship evaluation infrastructure that researchers depend on from day one.
Responsibilities:
- Run the full eval pipeline end to end and reproduce known results during onboarding, pairing with a senior engineer on your first task
- Build a judge calibration protocol: sample human-labeled decisions, measure agreement (κ, per-class P/R), identify drift zones, and document it so anyone can re-run it
- Extend an existing benchmark (GAIA, τ-Bench, SWE-bench slice, etc.) with new tasks targeting known capability gaps, including the prompt, environment, rubric, automated grader, and QA
- Run failure analysis on model outputs: categorize failure modes, quantify prevalence, and write up findings with recommendations for training data, judge prompts, or benchmark changes
- Own a recurring eval workflow (weekly regression suite, judge drift dashboard, red-team evaluation for a new capability) and ship tooling researchers actually use
Qualifications:
- 3+ years in software engineering, ML engineering, data science, or a research-adjacent role, with concrete evaluation experience from coursework, an internship, a side project, open source work, or a job
- Experience with at least one LLM evaluation framework (Harbor, Nemo Evaluator, etc.), with real opinions on what it does well and where it falls short
- Hands-on experience with LLMs: prompting, few-shot design, and ideally fine-tuning or RAG; regular use of coding agents
- Solid Python. You write clean, tested, version-controlled code that a colleague could run without you babysitting it
- Comfort with Git, CI/CD basics, Docker, and the Linux command line (SSH, tmux, debugging a remote job)
- Understanding of basic eval statistics: why accuracy misleads on imbalanced judges, what Cohen's κ measures, how to think about confidence intervals on a metric
- At least 3 of the following: you can explain why LLM-as-judge needs calibration; you've done failure analysis and can tell model bugs apart from prompt, grader, or retrieval issues; you know at least two agent benchmarks (GAIA, AgentBench, τ-Bench, MINT, SWE-bench, WebShop, ALFWorld) and a limitation of each; you've designed or extended an eval dataset with happy paths, edge cases, and adversarial examples; you've thought about non-determinism in eval, how you sample, how many runs, how you report variance
- You communicate clearly to both researchers and engineers, in the right language for each
- You're comfortable with ambiguity, can turn a half-formed request into a plan, and know when to ask for help
Preferred:
- RLVR / RLHF pipeline experience
- Training data curation experience
- Distributed eval orchestration experience
- Benchmark design from scratch
- Red teaming and adversarial eval experience
- Familiarity with psychometrics or measurement theory