Worked on enhancing stability in distributed machine learning pipelines by implementing FP8 rowwise padding for AllToAll pooled embeddings in the pytorch/torchrec and pytorch/FBGEMM repositories. Addressed evaluation-time crashes with small or irregular batch sizes by ensuring input lengths are divisible by the row dimension, applying zero-filled padding before communication and stripping it after decoding. Integrated this approach into both quantization and communication paths, maintaining model output consistency while improving CI and test reliability. Demonstrated expertise in Python, PyTorch, and distributed systems, with a focus on maintainable code and cross-repository consistency in quantization and data processing workflows.