Overview

This is a high-leverage, cross-functional role sitting at the intersection of ML research, data engineering, and product.

Responsibilities

• About the Role
• This is a high-leverage, cross-functional role sitting at the intersection of ML research, data engineering, and product. You'll own evaluation solutions end-to-end — from framework design through deployment and production monitoring — partnering closely with ML researchers, applied scientists, and product teams to define what "good" looks like for models and agents and turn measurements into actionable outcomes.
• Own evaluation solutions end-to-end: design the methodology, build the pipeline and tooling, deploy it into training and production workflows, and maintain/iterate on it over time — with minimal hand-off to other teams.
• Design and build evaluation frameworks and harnesses for LLMs and multi-step AI agents, covering offline benchmarks, online/production evals, and human-in-the-loop review.
• Define rubrics and scoring methodologies for open-ended tasks (e.g., tutoring quality, step-by-step reasoning, citation accuracy) where correctness isn't binary.
• Build automated, model-graded (LLM-as-judge) and rule-based evaluators, and validate them against human judgment for reliability and bias.
• Develop agent-specific evaluations: tool-use correctness, multi-turn task completion, planning/trajectory quality, failure recovery, and cost/latency tradeoffs.
• Create golden datasets, adversarial test sets, and regression suites that catch quality and safety regressions before they reach production.
• Partner with research teams to translate eval results into training signal — informing SFT/RLHF/RLAIF data curation, reward modeling, and fine-tuning priorities.
• Instrument production systems to collect real-world interaction data and feed it back into the evaluation and training loop.
• Build dashboards and reporting that give researchers, PMs, and leadership a clear, trustworthy view of model quality trends across releases.
• Drive eval methodology rigor: statistical significance, inter-rater reliability, sampling strategy, and avoiding metric gaming or overfitting to benchmarks.
• Collaborate with Trust & Safety and Legal/Compliance stakeholders to build evals for bias, hallucination, academic integrity, and other responsible-AI dimensions relevant to an education product.
• Mentor other engineers on eval best practices and help establish evaluation as a first-class part of the model development lifecycle.

Qualifications

• 5+ years of software/ML engineering experience, including hands-on work building or maintaining evaluation, testing, or measurement infrastructure for ML systems.
• Direct experience with foundation model evaluation and benchmarking beyond using APIs — experience gained at a foundation model lab or similarly frontier research environment is strongly preferred.
• Demonstrated ability to own an evaluation solution end-to-end — from initial design and dataset/methodology creation through pipeline build, deployment, and production monitoring — with minimal hand-off.
• Direct experience designing evals for LLMs and/or AI agents — e.g., benchmark design, LLM-as-judge pipelines, human annotation platforms, or A/B and offline/online eval frameworks.
• Strong programming skills in Python, PyTorch / Tensorflow and experience building production-grade data/ML pipelines.
• Hands-on knowledge of key model training and evaluation platforms, particularly AWS (e.g., SageMaker, Bedrock) and Databricks (e.g., MLflow, Unity Catalog, Delta Lake).
• Solid grounding in applied statistics — comfortable reasoning about sample size, variance, significance testing, and the limitations of aggregate metrics.
• Working knowledge of how LLMs are trained and adapted (pretraining, SFT, RLHF/RLAIF/DPO) and how eval signal feeds into that lifecycle.
• Familiarity with eval and observability tooling (e.g., internal or open-source frameworks for tracing, dataset versioning, experiment tracking).
• Excellent cross-functional collaboration skills; able to translate ambiguous product/research questions into concrete, measurable eval criteria.
• A bias toward rigor and skepticism — you instinctively question whether a metric is actually measuring what it claims to.

Nice to have

• Post-training & evals: hands-on experience with post-training techniques (SFT, RLHF, RLAIF, DPO) and the evaluation methodologies used to validate them, ideally gained at a frontier model labs
• Model benchmarking: experience building, running, or maintaining benchmark suites used to track and compare frontier model capabilities across training runs and releases.
• Data quality: experience with the data quality side of model training — curation, filtering, deduplication, and quality scoring of pretraining and post-training datasets.