Worked on the jeejeelee/vllm repository to deliver a modular weight-loading system and simplify the AXK1 model architecture. Developed the AXK1 AutoWeightsLoader, which streamlines the process of loading model weights, making the codebase more modular and maintainable. The approach involved removing redundant expert-management classes and methods, reducing architectural complexity and improving efficiency. These changes resulted in a cleaner, more extensible codebase that supports faster model deployments and easier future enhancements. The work was implemented using Python and PyTorch, leveraging deep learning and model optimization skills to improve maintainability and set a foundation for ongoing development in the project.