Deep Learning Solution Architect

NVIDIA NVIDIA · Semiconductors · Beijing, China

NVIDIA is seeking a Deep Learning Solution Architect to drive the research, development, and optimization of Reinforcement Learning algorithms and infrastructure for LLMs and multimodal models. The role involves collaborating with internal teams, improving customer engagements with NVIDIA RL technologies, and developing toolchains and documentation. Requires MS/PhD, 5+ years of experience in RL, LLM training, or multimodal learning, proficiency in PyTorch, and strong engineering skills in distributed training or orchestration.

What you'd actually do

  1. Drive research, development, and optimization of Reinforcement Learning algorithms and infrastructure for Large Language Models and multimodal models.
  2. Collaborate with internal research and engineering teams to adapt and validate state-of-the-art RL methods on NVIDIA GPU platforms at scale.
  3. Improve Reinforcement Learning initiatives and engagements with customers, providing technical guidance on integrating NVIDIA RL technologies into their AI workflows.
  4. Develop and maintain reusable toolchains, experiment management workflows, and technical documentation to accelerate both internal and customer-facing projects.

Skills

Required

  • MS or PhD in Computer Science, Artificial Intelligence, Mathematics, or related fields
  • solid foundations in algorithms and programming
  • 5+ years of experience (including research) in Reinforcement Learning, Large Language Model training, or multimodal learning
  • Proficient in PyTorch
  • familiar with RL training frameworks and workflows
  • Strong engineering skills
  • experience in distributed training, task orchestration, or evaluation pipelines
  • Ability to work independently with minimal day-to-day direction
  • willingness to conduct exploratory experiments on frontier problems
  • Outstanding verbal and written communication skills

Nice to have

  • Experience with RLHF, GRPO, DPO, or other alignment and post-training methods for LLMs
  • Experience with scale-out HPC or cloud architectures for large-scale model training
  • CUDA optimization or GPU performance tuning experience
  • Experience with agentic AI systems, code generation models, or multimodal RL
  • Publications in top-tier venues in RL, NLP, or multimodal learning

What the JD emphasized

  • Reinforcement Learning
  • Large Language Models
  • multimodal models
  • NVIDIA GPU platforms
  • RL training frameworks and workflows
  • distributed training
  • task orchestration
  • evaluation pipelines
  • exploratory experiments on frontier problems

Other signals

  • Reinforcement Learning algorithms
  • Large Language Models
  • multimodal models
  • NVIDIA GPU platforms
  • customer engagements
  • toolchains
  • experiment management workflows