Andrew Or worked on the pytorch/executorch repository, where he developed an internal QuantizedLinearWeightBase implementation to replace the previous reliance on torchao.quantization.subclass. By introducing this internal solution in Python using PyTorch, he decoupled executorch from an external quantization dependency, reducing maintenance complexity and external risk. To ensure a smooth transition and preserve existing API behavior, Andrew also added a legacy placeholder subclass for backward compatibility. His work focused on deep learning and machine learning infrastructure, demonstrating a thoughtful approach to dependency management and codebase stability, though the scope was limited to feature development without major bug fixes during this period.