Lead Research Engineer, Large-Scale Agentic RL Training
Core
Architect and optimize infrastructure for training frontier-scale agentic large language models (LLMs) with reinforcement learning, handling MoE models above 300B parameters and long contexts.
Role type
Lead Research Engineer (Large-Scale Agentic RL Training)
Builds
Asynchronous, fault-tolerant RL training pipelines and distributed training systems for agentic workloads.
Domain
Artificial Intelligence / Machine Learning Systems / Large Language Models
Deliverable
production ML models
Required skills
Large-scale model training, PyTorch, Megatron-Core, FSDP/DTensor, DeepSpeed, parallelism strategies (tensor/pipeline/context/expert), LLM inference engines (vLLM, SGLang), GPU profiling and optimization, numerical consistency, fault tolerance, technical leadership, mentorship, root-cause problem solving.
Preferred skills
RL post-training frameworks (NeMo-RL, verl, OpenRLHF, TRL), training models >100B parameters at 100k+ token contexts, distributed runtimes (Ray, Kubernetes, Slurm), numerical-parity testing, open-source contributions (vLLM, Megatron, PyTorch), publications at ML/systems venues.
Technologies
PyTorch, Megatron-Core, FSDP, DTensor, DeepSpeed, vLLM, SGLang, Ray, Kubernetes, Slurm, AWS, Azure, Google Cloud.
Responsibilities
Lead technical direction for the RL training stack architecture; optimize the full RL loop to reduce time per step and GPU idle time; design asynchronous pipelines for model-tool interaction; debug numerical, distributed, and performance issues; enable research scientists to test algorithms at scale; set engineering standards and mentor the team.
Seniority
Senior, hands-on IC with leadership responsibilities.