ML Platform Engineer
Preference Model
ML Platform Engineer
Preference Model
Reinforcement learning environments for frontier AI
Preference Model is a superintelligence data research company. We build RL environments for training capable, better-aligned superintelligences.
Almost every aspect of how a model behaves is shaped by the reward signals it's trained on. The biggest problem with AIs today is that they don't always do what we intend. Sometimes that's because they aren't capable enough, and sometimes it's because they aren't aligned. Both problems come down to the training objectives.
We're trying to address both of these issues in the most direct, highest leverage way available to us, which today means researching how to build RL environments better, not just for the AIs of today but for the AIs as they will be when they outsmart us.
Over the past year, we've built RL environments for several frontier labs, and we're backed by $16M in seed funding led by a16z. Our founding team has previous experience on Anthropic’s data team building data infrastructure, and datasets behind Claude.
An RL environment is only useful once it runs reliably inside a lab's training stack. Every environment we build has to be packaged, sandboxed, run at scale, checked against each lab's requirements, and delivered in the format that lab's platform expects. Much of that work is still manual today, and it takes time away from the people designing environments.
As an MTS on our ML Platform team, you will own the systems that take an environment from a builder's repo to a lab's training run. That includes Karotte, our open-source framework for defining RL environments, the internal platform that builds, runs, and reviews environments on CPU, GPU, and TPU, and the integrations that deliver them to each lab we work with. This is a senior ML platform engineering role where you will ensure what our Capabilities team builds runs and ships cleanly.
You'll join a small Platform team and work directly with environment builders and post-training researchers. In your first six months, success looks like a delivery path to each partner lab that is mostly automated: tasks marked ready get built, validated against that lab's requirements, and packaged without hand work. Our needs change quickly, so you'll also pick up other platform and infrastructure work as priorities shift.
Own the integrations that deliver our environments to partner labs: build container images, match each lab's submission format, and automate the checks that confirm requirements are met before anything ships. Do the delivery work by hand where you have to, then automate it.
Build and maintain Karotte, our open-source environment framework, from API design through releases and migrations that keep existing environments working.
Run and scale the infrastructure that builds, runs, and grades environments, including containerized rollouts and grading across CPU, GPU, and TPU, VM-based sandboxing, and the resource limits and isolation that stop a model from gaming its own grader.
Build the internal tools environment builders use every day (CLI, docs, data catalog, access controls, and run monitoring) so they spend their time on environment design instead of infrastructure.
Partner with environment builders and post-training researchers to turn recurring pain points into platform features, and ship quickly in response to their feedback.
5+ years of experience building and owning ML Platform Engineering, ideally AI related ML Platform Engineering
Strong software engineering fundamentals and a track record of building and operating production infrastructure or developer platforms that other engineers depend on.
Hands-on experience with containers and orchestration (Docker, Kubernetes, or similar) and at least one major cloud platform (GCP, AWS, or similar).
A working understanding of how to run untrusted code safely: isolation, resource limits, and what happens when a process fills the disk or refuses to exit.
You've owned something tedious end to end, such as a migration, a release process, or a delivery pipeline, and automated away as much of it as you could.
Extensive hands-on use of coding agents (Claude Code, Cursor, Codex, or similar), and an intuition for where models cut corners.
Proficiency in Python, including writing libraries and APIs that other people build on.
Experience building RL environments, eval harnesses, or agent benchmarks (Terminal-Bench, SWE-bench, or similar).
Familiarity with RL post-training and inference stacks (verl, Ray, SkyPilot, vLLM, SGLang, or similar).
You've built sandbox or code-execution infrastructure (microVMs such as Firecracker, gVisor, or similar) at an infrastructure startup or frontier lab.
You've been an early platform or infrastructure engineer at a startup and built the internal platform the rest of engineering relied on.
You've worked as a forward deployed or integrations engineer, getting your company's software running inside customers' own environments.
You've built developer tooling or internal platforms used by ML researchers, even outside a formal ML role.
You've maintained an open-source developer framework with real external users.
Competitive cash and equity compensation (>90th percentile)
Ownership and autonomy in a fast moving startup environment
Opportunity to work alongside senior and staff engineers from frontier labs and infrastructure companies, plus top ML engineers
Health, vision, dental, benefits
401K match
Lunch provided everyday onsite
Weekly snack orders
Visa sponsorship & relocation support available
We value diverse perspectives and experiences. If you're excited about this role but don't check every box, we still encourage you to apply.
Reinforcement learning environments for frontier AI
Preference Model builds reinforcement learning environments and harnesses to train frontier AI models, focusing on how reward signals shape model behavior. It designs robust RL harnesses and simulators that treat reward functions and graders as an attack surface, testing them across millions of training rollouts to harden against loopholes and gaming, with fail-safes for safety. Unlike others that only build models or tools, it targets the reward design and alignment bottlenecks, drawing talent from Anthropic, Stripe, Google DeepMind, and others, and backed by notable investors. Its goal is to broaden access to capable AI while preserving strong safeguards to keep frontier systems safe.
Explore more opportunities at Preference Model