Developed and delivered Tensor Direct Storage for the LMCache/LMCache repository, enabling connectors to store tensors directly without serialization. This refactor introduced new connector types for both bytes and tensors, updating the connector interface to support these data types and setting the default remote_serde to None for seamless end-to-end tensor storage. The work focused on backend development and system design using Python, targeting improved data throughput and reduced serialization overhead in distributed systems. By simplifying the connector surface and aligning with performance and scalability goals, the changes addressed the needs of tensor-heavy pipelines and enhanced the repository’s data handling capabilities.