Ary Gupta optimized the Top-K Top-P sampling kernel in the ROCm/aiter repository, focusing on improving throughput and reducing latency by restructuring block reduction logic outside of iterative loops. Using C++ and Python, Ary implemented comprehensive statistical and controlled tests to validate the accuracy and efficiency of the kernel, updating the test suite to use warnings for certain statistical checks to enhance reliability. Benchmarking tools were developed to quantify performance gains, and the codebase was refactored for better maintainability. Ary’s work demonstrated depth in GPU programming, performance optimization, and statistical testing, addressing both computational efficiency and code quality within a month.