Worked on the pytorch/torchrec repository to enhance the robustness of embedding lookups by addressing empty rank cases during inference. Developed a solution that introduces a dummy tensor path when no grouped configurations exist, preventing inference failures and stabilizing model serving. Enabled Dynamo tracing across the embeddings path, which improved observability and facilitated debugging for embedding-related workloads. The work focused on reducing production inference errors and aligning with performance monitoring goals. Utilized Python and PyTorch, applying skills in deep learning and data processing to ensure the embedding lookup code is more resilient and traceable in complex machine learning environments.