Worked on backend and infrastructure enhancements for ai-dynamo/dynamo and vllm-project/vllm-omni, focusing on stability, performance, and compatibility. Improved ai-dynamo/dynamo by preserving user-defined runtime device mappings in stage configuration and upgrading Docker images to v0.21.0, resulting in better resource allocation and faster, more reliable distributed runs. For vllm-project/vllm-omni, delivered a Diffusion LoRA Manager update to support SDXL pipelines, adding UNet component compatibility and robust layer replacement to streamline integration and reduce deployment friction. Employed Python, Docker, and containerization best practices, emphasizing maintainable code, repeatable deployments, and enhanced machine learning workflow reliability across both repositories.