Machine Learning Engineer — Reinforcement Learning

Key Responsibilities

• Distributed Framework Ownership – Extend or modify training frameworks (e.g., DeepSpeed, FSDP) to support new use cases and architectures.
• Training & Inference Integration – Connect training workers with rollout generation engines (e.g., vLLM, SGLang), including trajectory exchange and policy-weight synchronization.
• Launch Config & Debugging – Create and debug multi-node launch scripts with flexible batch sizes, parallelism strategies, and hardware targets.
• Metrics & Monitoring – Build systems for experiment tracking, job monitoring, and logging usable by collaborators and researchers.
• Infra Engineering – Write production-quality code and tests for ML infra in PyTorch or JAX; ensure reliability and maintainability at scale.
• RL Pipeline Development – Implement reward/verifier integration and trajectory processing, and validate log probabilities, token masks, and loss inputs with researchers.
• RL Execution & Recovery – Coordinate rollout and training workers, including checkpoint/restart and failure recovery; track policy versions and sample staleness when using asynchronous execution.

Must-Haves:

Nice-to-Haves:

Additional information

Qualifications

• 5+ years of experience in ML systems, infra, or distributed training
• Experience modifying distributed ML frameworks (e.g., DeepSpeed, FSDP, FairScale, Horovod)
• Strong software engineering fundamentals (Python, systems design, testing)
• Proven multi-node experience (e.g., Slurm, Kubernetes, Ray) and debugging skills (e.g., NCCL/GLOO)
• Ability to implement algorithms across GPUs/nodes based on mathematical specs
• Experience working on an ML platform/ infrastructure, and/or distributed inference optimization team
• Experience with large-scale machine learning workloads (strong ML fundamentals)
• Hands-on experience developing an LLM RL pipeline, with ownership across rollout/inference and distributed training integration.
• Working knowledge of policy optimization methods such as PPO or GRPO, sufficient to implement and debug sampling, log probabilities, and policy updates.
• Exposure to mixed-precision training (e.g., bf16, fp8) with accuracy validation
• Familiarity with performance profiling, kernel fusion, or memory optimization
• Open-source contributions or published research (MLSys, ICML, NeurIPS)
• CUDA or Triton kernel experience
• Experience with large-scale pre-training
• Experience building custom training pipelines at scale and modifying them for custom needs
• Deep familiarity with training infrastructure and performance tuning
• Experience with asynchronous RL, multi-turn or tool-using rollout environments, or custom reward/verifier systems.