Applied Research Scientist

About Fleet

We work with frontier labs, hyperscalers, and enterprises to build the environments, tasks, and evaluations that the next generation of agents are trained and judged on. We see this work, codifying human goals for agents and exposing where frontier models break, as the highest-leverage activity in the build-up to ASI. We've raised from Sequoia Capital, Menlo Ventures, BCV, and SV Angel, are growing very quickly, and are building the core infrastructure to unlock the next generation of agents.

About The Role

As an Applied Researcher, you will collaborate closely with our technical team to design and deliver robust environments, realistic data, and challenging tasks that are just out of reach of today's frontier models.

This role is also the bridge between our partner engagements and our core research team. You go deep on a lab engagement, come back with what you saw, and feed that into core research. When you find a thread worth pursuing, you pursue it.

Many projects will require herculean effort, attention to detail, and research taste. We want people who will do whatever it takes to solve immediate bottlenecks, and also think in systems and refuse local maxima. You will be at the frontier of AI development every day, on problems that, if solved, are among the biggest unlocks in front of us.

What You'll Do

• Understand the ask. Take a lab request and figure out what they actually need: not the literal spec, the underlying capability they're trying to teach or measure. Push back when the ask is wrong.
• Ship environments, tasks, and data. Deliver an artifact that runs end-to-end and is ready to go into a training or eval run.
• Be diligent. Know your failure modes. Your artifact will be used to train or evaluate a frontier model. A broken reward, a leaked answer, a degenerate solution path, a flaky environment can corrupt a multi-million-dollar run or invalidate a benchmark. Be the person who knows exactly how their artifact can break, and who has already closed the obvious ways.
• Close the loop. Each engagement teaches you something about how the lab's model fails, what reward shape actually trained, where an eval held up or didn't. That signal goes back to core research and shapes the next thread, recipe, or env primitive we build. Pursue your own threads when you find one worth pulling.
• Compound the platform. Reusable harnesses, eval protocols, dataset patterns, and infra primitives you build for one delivery become the substrate for the next.

What We're Looking For

We want people who can demonstrate they will work hard, learn fast, and go deep in at least one of: applied AI research, low-level infrastructure, or AI agent engineering. A former-founder track record, early-engineer-at-an-early-stage-startup background, or prior AI research experience is a strong plus.

• Strong engineer. Production Python; comfortable across containers, distributed systems, and a GPU when you have to.
• Deep understanding of ML. You know how models actually learn from environments: what makes a reward trainable, how RL training dynamics behave in practice, what supervision, preference, and RL data each do, where capability actually comes from.
• Research taste. You can look at a partner's request and tell whether the environment, the reward, or the data is what's going to determine whether the model learns. You can defend that view to a lab researcher.
• High ownership. You are the person who ships, and the person who owns the engagement end-to-end.
• Comfort with ambiguity. The frontier is, almost by definition, chaos. You should be energized by that, not paralyzed.
• Bonus: depth in a domain we want to teach a model.

How We Work

• Small, technical team.
• Speed with rigor.
• Truth over comfort. Honest with partners, with each other, and with ourselves.
• On-site in SF or NYC.

Compensation

Highly competitive salary and equity.