Research Scientist / Engineer Reinforcement Learning Infrastructure

present of AI; Eu; Hybrid; EUR 90,000 - 135,000; Full time

Job summary

Luma AI is seeking a Research Scientist / Engineer to design and optimize reinforcement learning infrastructure. You will enable training, rollout generation, environments, and reward checking on very large GPU fleets.

You should have hands-on experience scaling RL, building verifiers or agent environments, and solving asynchronous pipeline challenges. This role focuses on real impact beyond early-stage experiments.

Qualifications

• Experience with RL at scale and distributed systems.
• Ability to build environments and verifiers for agent tasks.
• Experience debugging asynchronous GPU pipelines.

Responsibilities

• Build the plumbing that lets reinforcement learning run on very large GPU fleets, from training and rollout generation to environments and reward checking.
• Address reward hacking and verification in large-scale, asynchronous pipelines.
• Work on post‑training RL with large language models and multi-modal models.

Job description

Research Scientist / Engineer Reinforcement Learning Infrastructure

The role is about building the plumbing that lets reinforcement learning run on very large GPU fleets, from training and rollout generation to environments and reward checking. A strong candidate has already post-trained large language models with RL, built verifiers or agent environments, and fixed stubborn problems in asynchronous pipelines. The posting is clear that real experience at scale matters, so this suits someone past the learning stage.

How it fits the AI era

RL infrastructure sits at the center of how modern models improve, so this work is closely tied to AI itself. AI tools can help write code and analyze logs, but the hard parts are still human: deciding what a reward should really measure, spotting reward hacking, and keeping huge runs stable. Judgment about what makes a model behavior useful stays with the people running the loop.

How it builds AI experience

Someone in this role would get hands‑on time with post‑training for multimodal models that handle video, image, audio, and text, which is still rare territory. Without that background, a good way to build toward it is to run small GRPO or RLVR experiments with veRL or TRL, use vLLM for rollouts, and build a sandboxed environment with a simple verifier. Showing a public project that handles multi‑step agent tasks would be especially relevant to what this company builds.

Worth showing

• RL post‑training at scale
• Distributed PyTorch parallelism
• vLLM or SGLang rollouts
• Agent environments and verifiers
• Debugging async GPU pipelines

This is our own read, written from the employer's posting and the company's Spotlight. The employer's posting is the source of truth for the role.

Luma AI builds multimodal AI models and agents for video, image, audio, and text creation, per the company.