GPU Performance Engineer — Scale Large AI Models & Training

Reka; United States; Remote; USD 150,000 - 210,000; Full time

Benefits offered by this job

Five weeks of paid leave

Comprehensive healthcare benefits incl

Visa assistance including H1B and OPT

Job summary

Reka is seeking an experienced ML/AI engineer to advance training infrastructure and optimize model performance at scale within a globally distributed team.

The role focuses on post-training processes (reinforcement learning and fine-tuning), improving GPU-accelerated workloads, and making serving infrastructure more efficient and scalable.

Applicants should be proficient in Python and PyTorch, with deep experience in CUDA/C++, scalable compute clusters, and profiling for bottlenecks.

Qualifications

• Strong engineering skills with fluency in Python and PyTorch.
• Proven experience implementing and training large deep learning models.
• Experience writing and debugging low-level GPU code (CUDA, C++).
• Experience scaling GPU jobs using large-scale compute clusters (Slurm or Kubernetes).
• Ability to analyze and optimize GPU-accelerated workloads through profiling and bottleneck identification.

Responsibilities

• Design and implement improvements to training infrastructure.
• Contribute to technical decisions to optimize model performance.
• Work on post-training processes including reinforcement learning and fine-tuning.
• Improve the efficiency and scalability of model serving infrastructure.

Skills

Python

PyTorch

CUDA

C++

GPU Compute

Tools

Slurm

Kubernetes

Job description

Reka is seeking an experienced ML/AI engineer to advance training infrastructure and optimize model performance at scale within a globally distributed team.

The role focuses on post-training processes (reinforcement learning and fine-tuning), improving GPU-accelerated workloads, and making serving infrastructure more efficient and scalable.

Applicants should be proficient in Python and PyTorch, with deep experience in CUDA/C++, scalable compute clusters, and profiling for bottlenecks.