RL Systems Architect: Scale Reinforcement Learning Across GPUs

Luma; New York (NY); On-site; USD 195,000 - 395,000; Full time

Job summary

Luma is building the systems that drive reinforcement learning at frontier scale, coupling policy optimization with thousands of GPUs and large inference fleets. You’ll own end-to-end RL post-training stacks, from trainer to rollout to reward and verification, aimed at reliability and speed.

We seek experts who have deployed RL at scale, with experience in post-training LLMs, distributed PyTorch, and sandboxed environments.

Qualifications

• Hands-on experience post-training LLMs with RL at scale.
• Extensive distributed PyTorch training and parallelism for foundation models.
• Experience building RL environments, reward functions, verifiers, or evaluation harnesses for LLM agents.
• Deep familiarity with RL post-training frameworks and rollout inference engines (veRL, OpenRLHF, TRL, Ray).
• Strong understanding of GPU clusters, networking, and communication libraries (NCCL, MPI) under mixed workloads.

Responsibilities

• Design, build, and scale distributed RL post-training systems across thousands of GPUs.
• Build high-throughput rollout generation, integrating inference engines (vLLM, SGLang) and synchronization.
• Design RL environments for multi-step tasks with sandboxed execution and tool use.
• Build reward infrastructure and RL judge pipelines, defending against reward hacking.
• Develop evaluation, monitoring, and debugging tooling for large RL runs.
• Advance training efficiency and translate ideas into production with researchers.

Skills

Post-training LLMs with RL

Distributed PyTorch

RL environments & rewards

RL frameworks & inference engines

GPU clusters & NCCL/MPI

Tools

vLLM

SGLang

NCCL

MPI

Ray

Job description

Luma is building the systems that drive reinforcement learning at frontier scale, coupling policy optimization with thousands of GPUs and large inference fleets. You’ll own end-to-end RL post-training stacks, from trainer to rollout to reward and verification, aimed at reliability and speed.

We seek experts who have deployed RL at scale, with experience in post-training LLMs, distributed PyTorch, and sandboxed environments.