Research Scientist / Engineer – Performance Optimization

Role overview This position centers on making large multimodal models fast at every layer of the stack, from training through production deployment. The work combines low-level GPU optimization with architectural tuning for transformer-based systems, ensuring models run efficiently on modern accelerators without sacrificing output quality. It is a deeply technical role for someone who lives in profilers, kernels, and hardware-aware design.

Responsibilities - Profile and optimize GPU, CPU, and accelerator code paths to push utilization higher and latency lower. - Write high-performance PyTorch, Triton, and CUDA, including custom operations where off-the-shelf kernels fall short. - Design and ship fused kernels that take advantage of tensor cores and platform-specific hardware features. - Tune model architectures and implementations for distributed, multi-node production deployment. - Build the monitoring, analysis, and automation tooling that prevents performance regressions over time. - Research and apply state-of-the-art optimization techniques targeting transformer internals.

Requirements - Expert-level proficiency in Triton and CUDA, with a track record of serious GPU optimization work. - Strong PyTorch skills, including kernel development and custom-operation authoring. - Hands-on experience with profiling tools such as NVIDIA Nsight, the PyTorch profiler, and bespoke instrumentation. - Deep understanding of transformer architectures and attention mechanisms.

Nice to have - Familiarity with compilers and exporters such as torch.compile, TensorRT, ONNX, or XLA. - Experience tuning inference workloads for latency and throughput under real-world traffic. - Deep knowledge of the Triton compiler and kernel fusion techniques. - Comfort with warp-level intrinsics and advanced CUDA optimization patterns.

Benefits and work setup The role is positioned as an equal-opportunity position with the employer described as an equal opportunity employer; specific compensation, benefits, and remote/hybrid policies were not provided in the source material.