Worked on the pytorch/torchrec repository to enhance the reliability and performance of embedding operations in distributed deep learning systems. Focused on stabilizing the VBE+SSD path by addressing a runtime error in the StagedPipeline, ensuring the correct batch size was passed to SSDTableBatchedEmbeddingBags and resolving tensor splitting failures. Additionally, refactored the prefetch pipeline for GroupedPooledEmbeddingsLookup to guarantee accurate propagation of VBE metadata, which improved both reliability and potential performance of embedding lookups. Leveraged Python and PyTorch, applying expertise in data processing and distributed systems to deliver targeted improvements that addressed both feature development and bug resolution.