Worked on the vllm-omni repository to deliver a modular enhancement for text-to-speech workflows by refactoring the Qwen3-TTS codebase. Developed a reusable prompt-embeds builder and introduced a persistent tts_pad_embed buffer, optimizing memory usage and reducing allocation overhead during audio processing. The approach focused on isolating core logic for improved maintainability and future extensibility, enabling easier testing and feature iteration. Collaborated effectively within the codebase, as reflected in co-authored commits, and maintained clear ownership of changes. Utilized Python and applied full stack development and machine learning skills to achieve a focused, feature-driven iteration without major bug fixes during the period.