Worked on GPU compilation pipelines in the Intel-tensorflow/tensorflow and openxla/xla repositories, focusing on autotuner pass placement and offline autotuning infrastructure. Developed configurable debug options in C++ to control GEMM and Conv autotuning pass ordering, enabling performance experimentation and standardization across TensorFlow and XLA. Enhanced backend flexibility by introducing version-aware cache keys and protobuf-based schemas for offline autotuning, while refactoring code to remove unused experimental cache logic. Leveraged skills in compiler design, performance optimization, and system architecture to streamline backend workflows, reduce remote tuning dependencies, and lay a scalable foundation for future performance improvements in GPU workloads.