GPU Kernel Developer for AI Model Evaluation

Review GPU and accelerator kernel development tasks that support the training and evaluation of advanced AI models. This role focuses on assessing task quality, numerical correctness, completeness, fair performance benchmarking, appropriate scope, and whether kernels compile and run successfully across varied task types. You will provide clear written feedback using defined evaluation rubrics. Key Responsibilities

• Evaluate GPU and accelerator kernel development tasks for quality, correctness, completeness, scope, compilation validity, and runtime validity.
• Assess numerical accuracy, including tolerance selection and reference implementation choices.
• Review performance benchmarks for fairness and sound methodology.
• Deliver clear, rubric-based written feedback across diverse kernel task categories.

• At least 3 years of hands-on experience developing, optimizing, or verifying GPU or accelerator kernels in at least two of the following: CUDA, Triton, NKI, or Pallas with JAX.
• Strong knowledge of kernel numerical-correctness standards, including absolute, relative, and ULP tolerances and reference implementation selection.
• Experience profiling and benchmarking performance with tools such as Nsight, NCU, roofline analysis, or framework-native profilers.
• Familiarity with compilation and runtime issues, including driver mismatches, out-of-memory errors, launch configuration errors, shape or stride mismatches, and autotuning failures.
• Experience with at least three kernel task types: specification-based generation, cross-framework translation or lowering, hardware-target migration, debugging, performance optimization, or operator fusion.

• Experience spanning NVIDIA GPU environments, including CUDA or Triton, and custom accelerator environments, including NKI, Pallas, or TPU.
• Background in compiler engineering, MLIR, or intermediate-representation lowering.
• Knowledge of memory-hierarchy optimization, including shared-memory tiling, register pressure, bank conflicts, and coalesced memory-access patterns.
• Contributions to kernel libraries such as cuBLAS, cuDNN, Triton community kernels, or JAX/XLA custom calls.

• Remote role open to candidates located in the United States.
• Hourly engagement.

• $70 to $90 per hour.