Drive down wall-clock time to convergence by profiling and eliminating bottlenecks across the foundation model training stack stack, from data pipelines to GPU kernels
Design, build, and optimize distributed training systems (PyTorch) for multi-node GPU clusters, ensuring scalability, robustness, and high utilization
Implement efficient low-level code (CUDA, cuDNN, Triton, custom kernels) and integrate it seamlessly into high-level training frameworks
Optimize workloads for hardware efficiency: CPU/GPU compute balance, memory management, data throughput, and networking
Develop monitoring and debugging tools for large-scale runs, enabling rapid diagnosis of performance regressions and failures
Deep experience in distributed systems, ML infrastructure, or high-performance computing (8+ years)
Production-grade expertise in Python
Low-level performance mastery: CUDA/cuDNN/Triton, CPU–GPU interactions, data movement, and kernel optimization
Scaling at the frontier: experience with PyTorch and training jobs using data, context, pipeline, and model parallelism
System-level mindset with a track record of tuning hardware–software interactions for maximum utilization