Worked on the vllm-project/vllm-omni repository to establish the foundational architecture for multimodal output handling, enabling the system to process and extend support for text, image, and audio data. Focused on implementing phase one of decoupled output types, the work introduced structured payloads and output modalities, laying the groundwork for scalable and modular feature expansion. Leveraged Python programming and data structures to create a flexible framework that supports diverse data types and simplifies future pipeline extensions. Emphasized robust unit testing and multimodal processing techniques to ensure maintainability and adaptability as the project evolves to accommodate new requirements.