Worked on the nvidia-cosmos/cosmos-transfer1 repository to enhance the process of converting Tensor Parallel checkpoints to Fully Sharded Data Parallel (FSDP) format. Developed a robust Python script with argument parsing and an option to include the base model, streamlining large-model training deployments. Improved documentation using Markdown, providing clearer usage instructions and example commands to reduce onboarding friction. Addressed critical issues in the TP to FSDP conversion logic, validating integration with large-scale training workflows. These updates improved reproducibility and accelerated FSDP adoption in production environments, demonstrating depth in deep learning, scripting, and model conversion within a focused one-month development period.