Worked on the elixir-nx/emlx repository to deliver a device-aware tensor copy optimization focused on improving backend performance and device interoperability. Developed a new copy path that minimizes unnecessary host round-trips when copying tensors within the same device, addressing both standard and quantized tensor types. Enhanced backend copy functionality to ensure independent tensor references are maintained during operations, supporting robust data handling. Expanded test coverage to validate copy behavior across various devices and tensor types, including quantized tensors. Utilized C++, Elixir, and backend development skills to align the work with project goals, laying the foundation for measurable performance improvements and maintainability.