Research Engineer, Post-Training

Strong judgment about model behavior: you can read traces, inspect outputs, identify failure modes, and reason about whether a metric is measuring the thing that matters.

Strong Python and research-engineering ability.

You can write clean code, debug experiments, and build the simple but reliable systems needed to make research move faster.

Experience building data or evaluation infrastructure for ML workflows, such as dataset curation pipelines, model-output processing, experiment tracking, evaluation dashboards, or regression analysis tooling.

Experience with distributed training, inference systems, GPU workloads, or large-scale ML experimentation.

Research publications, open-source contributions, or shipped industry work in LLMs, agents, evaluation, or ML systems.

$231,000 - $340,000

Depending on your location, an Applicant Privacy Notice may apply to you. You can find all of our Applicant Privacy Notices [

here

].

#LI-AK1

Harvey is an equal opportunity employer and does not discriminate on the basis of race, gender, sexual orientation, gender identity/expression, national origin, disability, age, genetic information, veteran status, marital status, pregnancy or related condition, or any other basis protected by law.

Post-training is how Harvey turns expert feedback and agent traces into models that are meaningfully better at legal work.

We are looking for a research engineer who can help scale that loop: defining and running model training experiments, interpreting results, and working with internal and external research partners to build better data, environments, graders, and training recipes.

The ideal candidate has extensive hands-on experience training open weight models, either in a research or production setting, and enough engineering depth to run and debug experiments efficiently.

Drive post-training experiments, pushing agent performance while navigating the Pareto frontier of cost, latency, security, and governance.

Optimize agent harnesses, including domain-specific skills, tools, subagents, retrieval strategies, and validation loops that improve quality on long-horizon legal work.

Study agent behavior, identifying patterns that correlate with successful work product, and converting those findings into training data, evals, or harness changes.