Worked on performance and data processing optimizations across ROCm/FBGEMM and pytorch/torchrec repositories, focusing on embedding pipelines. Delivered a faster and more robust initialization path for SSDTableBatchedEmbeddingBags in ROCm/FBGEMM by disabling compaction during bulk initialization and improving memory safety through byte-based chunk sizing, addressing out-of-memory risks with variable row dimensions. Added targeted test coverage to validate these changes. In pytorch/torchrec, introduced a dedicated prefetching class to the customized order pipeline, enabling more efficient data readiness for computation. Leveraged C++, Python, and unit testing to ensure reliability and maintainability, with an emphasis on memory management and database optimization.