Arbaz worked on the pytorch/FBGEMM repository, focusing on improving the reliability of jagged tensor operations by addressing a memory synchronization bug in the Jagged Index Select path. Using C++ and CUDA, Arbaz identified and resolved a crash caused by shared memory being overwritten before all threads completed their reads, which previously led to instability in multi-threaded GPU execution. The fix ensured proper synchronization, reducing the risk of memory violations in production workloads. Benchmarks across NVIDIA and AMD platforms confirmed stable performance with no regressions. Arbaz’s work demonstrated depth in GPU programming and parallel computing, emphasizing robust, maintainable code changes.