Shangz worked on core stability and correctness improvements in both the ROCm/jax and NVIDIA/TransformerEngine repositories, focusing on low-level programming and performance optimization using C++ and Python. In ROCm/jax, Shangz addressed intermittent errors in distributed tensor workloads by refining the squeeze lowering rule to preserve sharding information during tensor shape transformations, ensuring more reliable sharding propagation. In NVIDIA/TransformerEngine, Shangz enhanced large-scale tensor handling by widening the numel() return type from int to size_t, preventing overflow and improving memory estimation for massive workloads. The work demonstrated careful attention to type safety and robust tensor manipulation in high-performance computing environments.