In February 2026, Johannesz contributed a major performance optimization to the pytorch/pytorch repository by reworking the eigenvalue computation backend for torch.linalg.eigh. He enabled unconditional dispatch to the linalg_eigh_cusolver_syevj_batched backend, leveraging CUDA and C++ to achieve over 100x speedups on relevant workloads. By removing the unused syevj backend and updating CUDA heuristics, Johannesz streamlined the codebase and reduced maintenance complexity. His work focused on linear algebra libraries and performance optimization, resulting in faster experimentation and more scalable model training. The depth of this contribution reflects strong expertise in both low-level CUDA programming and high-level numerical computing.