Worked on the mirage-project/mirage repository, focusing on enhancing the Llama3 demo by introducing a command-line argument to control Cutlass kernel usage. This addition allowed users to select kernels at runtime, streamlining experimentation and benchmarking processes. The implementation involved Python-based command-line interface development, ensuring the new argument was properly integrated and propagated throughout the demo. By addressing the missing use_cutlass_kernel argument, the work improved the demo’s configurability and reduced setup time for internal tests. The changes supported more reproducible benchmarks and flexible kernel selection, reflecting a methodical approach to feature delivery and attention to runtime configuration in Python environments.