Founding Engineer, RL Environments + Data

Doe’s founders are bottlenecked by the work required to deliver data and RL contracts while continuing to improve the product. The company needs an engineer who can turn the large volume of agent traces it generates into reliable environments, benchmarks, and evaluation data, freeing the founders to ship product and scale the business.

You’ll investigate where models fail, select and quantify meaningful failures, and package them into environments that reflect real business work. The primary spike is building simulated company environments—complete with software such as inboxes, CRMs, records, and files—that can evaluate how agents perform over extended workflows and produce benchmarks whose rankings correlate with production performance.

What You'll Own

• Build synthetic environments that model the software and workflows agents use to perform business tasks.
• Develop graders and evaluation systems that remain reliable when agents exploit weaknesses or reward-hack the evaluation.
• Create benchmarks whose rankings align with agent performance in production.
• Review and triage agent traces to identify meaningful, previously unquantified model failure modes.
• Package discovered failure modes into environments and datasets for model evaluation, product improvement, and potential delivery to AI labs.
• Work with the founders on data and RL contracts while improving the evaluation harness and overall agent experience.

Technical Requirements

• Experience with RL environments, synthetic data, post-training, benchmark development, dataset building, or model evaluation.
• Ability to determine whether a dataset, trace, or model failure is meaningful and valuable.
• Experience building or contributing to environments, benchmarks, or datasets used to train or evaluate models.
• Ability to simulate software such as an inbox or CRM well enough to withstand agent tool calls.
• Ability to reason about graders, evaluation design, model failure modes, and reward hacking.
• This is an environments and data role rather than a product-facing engineering or pure research role.

Execution and Ownership

• Own ambiguous evaluation and environment-building problems from identifying the failure mode through packaging and delivery.
• Explain technical decisions, trade-offs, and the work represented on your resume in a technical interview.
• Work autonomously without hour tracking or close day-to-day management; output is prioritized over a fixed schedule.
• Be willing to test new tools and working methods, especially as AI workflows evolve.
• Provide specific feedback and communicate clearly with the founders and relevant stakeholders.

Background

• Typically one to seven years of experience, with at least one year out of school preferred.
• Strong data-lab experience is relevant, particularly work involving synthetic data or post-training; human-expert work alone is not sufficient.
• Experience at organizations such as DataCurve, Mercor, Surge, or AfterQuery may be relevant when the work involved synthetic data, post-training, or model evaluation.
• A founder background, founder fellowship, or standout achievement in a niche area can be a positive signal, but no rigid school or pedigree requirement is specified.
• Candidates should be able to demonstrate a concrete example of building or contributing to a benchmark, dataset, environment, or model-evaluation system.

Nice to Have

• Experience finding a real flaw in a widely used benchmark.
• Experience catching a model reward-hacking a grader and fixing the grader.
• Experience building environments or benchmarks used by a frontier lab for training or evaluation.
• A background as a founder or participation in YC, SPC, Thiel, or a similar fellowship.
• High-level competition in mathematics, programming, debate, or games.
• A general understanding of how benchmarks and datasets can create demand or support commercial model evaluation.

Who Will Thrive Here You’ll thrive here if you enjoy working at the frontier of model evaluation, can turn ambiguous failure modes into concrete environments and benchmarks, and are comfortable owning work end to end. The team values people who are kind and easy to work with, willing to try new tools, and able to operate independently in a fast-moving environment.