About the Role:

The bottleneck in building reliable, useful AI is no longer writing code. It's defining what good looks like and measuring how well our products deliver it. That takes high-quality data and evals that teams can trust.

Our team owns data and evaluation science for Glean's AI products. We build evalsets, agentic LLM-judges, and benchmarks that capture what good looks like, along with the platform and infrastructure that run evals reliably at scale, and enterprise simulation environments where we test how our assistants and agents actually perform. Starting from real production usage, you'll turn what we learn into evals that guide improvements across the entire stack, from agent harnesses to tool use, skill iteration, model routing, and post-training of task specific models.

You'll bring new research into production and partner with R&D, product, and GTM to turn findings from evals into better products and customer outcomes. If you care about quality and want to build the systems that measure it, this role is for you.

You will:

• Build evalsets and benchmarks from expert-curated examples, production data, synthetic data, and past failures, covering edge cases and key user slices.
• Build and scale the eval platform and agent observability: experiment pipelines, trace capture and replay, and dashboards, with reproducible, affordable runs.
• Build agentic LLM verifiers for correctness, groundedness, citation quality, and task success, and calibrate them against human labels using clear rubrics.
• Run evals at every stage, from CI and pre-release suites to sampled production traffic, and report results with confidence intervals.
• Build enterprise simulation environments with simulated users and realistic company data to test multi-step agent tasks.
• Diagnose failures by slice and trace, and turn eval results into concrete fixes, weighing quality against latency and cost.

About you:

• 5+ years of software engineering experience with strong Go and/or Python and backend fundamentals: APIs, data modeling, testing, and distributed pipelines. TypeScript is a plus.
• Experience building datasets and benchmarks for LLM use-cases, RAG, or agents, including evalset curation and synthetic data generation.
• Familiarity with eval methodologies and their tradeoffs: deterministic metrics, LLM-as-a-judge, pointwise vs pairwise grading, human labeling, and judge calibration.
• Experience building eval or experiment infrastructure that runs reliably at scale, including CI/CD integrations and observability tooling like OpenTelemetry.
• Strong analytical skills: you dig into traces and failure cases, write SQL, and use data and metrics to explain why quality moved.
• Nice to have: experience with simulation environments and/or model post-training techniques.

• This role is hybrid (4 days a week in our Mountain View or San Francisco office)