During March 2026, Subhadeep Chakraborty focused on improving the correctness and stability of gradient computation for mixed-dtype inputs in the MSELoss function within the pytorch/pytorch repository. He addressed a bug where gradients were not properly computed for tensors with different data types by implementing targeted fixes in the backward pass. Using C++ and PyTorch, he enabled input promotion to a common dtype and enforced safe casting in the TensorIterator, ensuring gradients accurately reflected input types. This work enhanced numerical reliability and training stability for mixed-precision deep learning models, demonstrating a strong understanding of both machine learning and low-level systems programming.