Strong generalist engineer with full-stack capabilities.
Proficient in TypeScript/Python or similar languages.
Ability to work with ambiguous requirements and translate them into concrete deliverables.
Excellent communication skills with the ability to engage customers and researchers effectively.
Genuine curiosity about evaluation metrics and agent performance.
Responsibilities
Transform research hunches into actionable and measurable environments.
Develop high-quality environments, task distributions, and datasets under tight schedules.
Recognize and generalize one-off builds into core products for broader impact.
Collaborate closely with researchers to ensure product alignment with actual needs.
Iterate on deliverables based on customer feedback and performance metrics.
Benefits
Work alongside a highly skilled team from renowned tech companies.
Engage with cutting-edge technologies in AI and data.
Direct involvement with real customers and tangible projects.
Ownership of complete deliverables and customer relationships.
Full Job Description
What You'd Work On
The hardest problems at the frontier. A new environment modality, an eval targeting a failure mode nobody's measured, a dataset that doesn't exist yet. You take problems like these from a researcher's hunch to a shipped deliverable, working at the edge of what agents can currently do.
Environments, evals, and datasets. One project is a high-fidelity environment, the next is a task distribution with grading logic, the next is a dataset built to a demanding spec. The bar is frontier-lab quality and the pace is relentless - you're writing whatever the deliverable needs: environment code, task specs, scoring harnesses.
Pulling the frontier into the platform. The best one-offs don't stay one-offs. You'd recognize when a custom build proves out a capability worth generalizing, and help fold it into the core product - so it compounds instead of sitting on a shelf.
About You
Full-stack range. You're a strong generalist engineer - comfortable across backend services, data pipelines, and enough frontend to ship a usable interface. TypeScript/Python or similar. You'd rather own a whole deliverable than a layer of one.
Fast in ambiguity. Requirements arrive as a hunch, not a spec. You can turn a loosely-defined research question into a concrete environment, task set, eval, or dataset - and validate that it measures what was meant, not just what was easy to build.
Customer instincts. You communicate clearly with researchers and technical customers, push back when they're asking for the wrong thing, and know the difference between what someone requests and what they need. The work moves between deep async stretches and high-bandwidth sessions with customers, in person when it counts.
Evals curiosity. You don't need an ML research background, but you're genuinely interested in how agents fail and how to measure it. You'll learn task design, LLM-judged scoring, and reward hacking detection on the job.
Experience. No hard rule. Ideally at least 3 years shipping production software, but less works if the above sounds like you.
What Makes This Different
You see the frontier first. The problems that land on your desk are the ones frontier researchers haven't solved yet. What you build is often the first working version of something the field will need - and you're the one who proves it's possible.
Real customers, real deadlines. Every project has a named customer and a researcher waiting on the output. This isn't building a platform and hoping someone uses it.
Ownership, not theater. You own whole deliverables and customer relationships, not tickets in a queue. One week you're shipping a custom eval for a lab, the next you're generalizing it into the core product.
The team. Our team is ex-Stripe, Snap, AWS, Microsoft, Airtable - you'll work with a small team who has years of shipping high-impact products over the last decade.
Cutting edge. You will get to touch the latest and greatest technologies across the data, AI, and infrastructure stack.
