Senior Software Engineer - AI Evals

Job Description About the Role

Chegg is building the next generation AI Evaluation Platform to keep improving learner experience across product and services from copilots to autonomous agents. As we scale our use of large language models and agentic systems, the quality, safety, and reliability of these systems depend on rigorous, well-designed evaluations. We’re hiring a Senior Engineer to own the design and implementation of evaluation frameworks (“evals”) that measure model and agent performance across accuracy, reasoning, safety, and pedagogical quality — and that directly inform how we train, fine-tune, and ship AI systems.

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.

What You’ll Do

• 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.

What We’re Looking For

• 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.
• Experience with agentic architectures — tool calling, multi-step planning, memory, orchest…