AI Inference & GPU Workload Optimization Lead

STN, Inc.; United States; Remote; USD 180,000 - 280,000; Full time

Job summary

STN, Inc. is seeking an Inference & Workload Optimization Lead (Principal) to own profiling of customer AI workloads and drive scalable optimizations across clusters.

The role focuses on data loading bottlenecks, memory bandwidth, and kernel efficiency to accelerate training and inference on dedicated GPU capacity. The candidate will establish performance baselines per hardware generation, starting with current-generation NVIDIA hardware, and translate expertise into agent-executable

Qualifications

• Deep GPU performance engineering: profiling memory bandwidth and kernel efficiency.
• Proficient with kernel-level optimizations, batching and data layout for large-scale workloads.
• Experience with distributed training and inference at scale, including NCCL/tensor parallelism.

Responsibilities

• Profile customer training and inference workloads to identify bottlenecks such as data loading and memory bandwidth.
• Establish hardware-specific performance baselines per generation starting with NVIDIA GPUs.
• Convert expertise into agent-executable optimization so fixes apply across clusters.
• Collaborate with operating teams and customers on workload tuning and deployment.
• Own internal model serving as a secondary responsibility including cost- and latency-aware routing.
• Advise on GPU allocation between internal inference and customer workloads.

Skills

GPU performance engineering

Profiling tooling fluency

Explain findings to customers

Agent-embedded optimizations

Tools

Nsight

DCGM

Framework profilers

Job description

STN, Inc. is seeking an Inference & Workload Optimization Lead (Principal) to own profiling of customer AI workloads and drive scalable optimizations across clusters.

The role focuses on data loading bottlenecks, memory bandwidth, and kernel efficiency to accelerate training and inference on dedicated GPU capacity. The candidate will establish performance baselines per hardware generation, starting with current-generation NVIDIA hardware, and translate expertise into agent-executable