During May 2026, contributed to the pytorch/torchrec repository by implementing total gradient norm tracking within the GradientClippingOptimizer. This feature exposed the last_total_grad_norm property, allowing users to monitor pre-clipping gradient norms and clipping frequency across NORM, VALUE, and NONE modes. The approach involved persistent tracking across optimizer steps, enabling seamless integration with TensorBoard for enhanced training diagnostics without altering training semantics. Leveraging Python, PyTorch, and gradient optimization techniques, the work improved observability and stability monitoring for MVAI training pipelines. The implementation aligned with existing clipping mechanisms and underwent thorough code review, reflecting a focused and technically sound engineering contribution.