Research Engineer, RL Environments & Synthetic Data (all genders)

ellamind • Berlin, KY

4d • Remote

About The Position

At ellamind, the Research Engineer, RL Environments & Synthetic Data builds realistic, verifiable training environments and synthetic data for language models. The role works directly with model-training researchers, observes how tasks affect model performance, and contributes to decisions about what models should learn next.

The work includes creating environments for long, multi-step tasks involving tools, documents, and code; building generators that produce diverse tasks at varying difficulty; developing reliable tests and judges; identifying and preventing model shortcuts; and measuring the effects of task sets on models. The role also helps decide which environments, datasets, and benchmarks to release publicly, with the first open-source release planned for Q1 2027.

The position is based remotely anywhere in the EU, ideally Germany, with optional offices in Bremen or Berlin. The company develops language models and methods for improving them, and shares research and resources with the open-source community.

Requirements

• Experience building something used to train or evaluate an LLM, such as an environment, benchmark, evaluation harness, or data pipeline.
• Strong Python skills and sound engineering practices.
• Comfort working with containers, sandboxes, and asynchronous code.
• Ability to inspect actual data samples and model trajectories and identify problems.
• Understanding of what makes a task realistic, verifiable, and difficult to game.
• Clear written communication sufficient to make results understandable and trustworthy.
• Live in the EU and be authorized to work there.
• English proficiency.

Nice To Haves

• Experience with reinforcement learning, including GRPO- or PPO-style training or reward design.
• Knowledge of a real-world work domain such as law, finance, insurance, or public administration.
• Experience with sandboxing infrastructure such as Docker, Firecracker, or Kubernetes.
• Open-source contributions.
• German-language skills or familiarity with German-language processes.

Responsibilities

• Build environments for long, multi-step work involving tools, documents, and code.
• Develop generators that scale task ideas into large, varied task sets, including difficult tasks, with controls for difficulty and checks for duplicates or leaked answers.
• Create and maintain verification methods such as tests, state checks, and rubric-based judges.
• Review model rollouts to identify shortcuts and cheating, close those loopholes, and build quality-assurance processes.
• Measure how task sets affect model performance and recommend which tasks or environments to scale or discontinue.
• Contribute to decisions about publicly releasing environments, datasets, and benchmarks.
• Work directly with the training team and use observed training results to inform future tasks.

Benefits

• Work remotely from anywhere in the EU, ideally Germany.
• Access to hundreds of GPUs across EuroHPC and commercial compute.
• Use AI tools that make the work easier.
• Attribution on public work.
• Access to founders and a research team.
• Optional offices in Bremen or Berlin; the Bremen office has free drinks and smoothies and a boat.
• Competitive pay and equity options.

What This Job Offers

• Job Type: Full-time
• Career Level: Mid Level
• Education Level: No Education Listed