Worked on reliability and benchmarking improvements for machine learning workflows in the pytorch-labs/helion and pytorch-labs/tritonbench repositories. Focused on enhancing accuracy testing and error handling, the developer implemented safeguards to prevent unintended code output during eager reference mode and expanded benchmarking capabilities by introducing an addmm operation with scalar bias. In tritonbench, addressed accuracy-critical issues by ensuring proper weight copying for fair operator comparisons. Leveraged Python, PyTorch, and Pytest to deliver these changes, emphasizing robust testing and model optimization. The work contributed to more trustworthy performance signals and fairer evaluation of deep learning operator implementations across both repositories.