Worked on the vllm-project/vllm-omni repository to migrate the MammothModa2-Preview example to a standardized task format, focusing on improving the text-to-image generation workflow. Leveraged Python for development, integrating shared offline image resources and model extras to streamline asset reuse and enable faster experimentation. Emphasized model deployment and image processing skills to ensure the new format supported both production-readiness and easier onboarding for future contributors. Collaborated with team members through Git-based workflows, maintaining clear code ownership and documentation. The work standardized task execution across the project, enhancing maintainability and supporting more efficient iteration for machine learning model development.