Developed and delivered a video processing feature for the jeejeelee/vllm repository, focusing on temporal compression with tubelet grouping for the Nemotron-3-VL model. This approach grouped video frames into tubelets, enabling the model to process longer video sequences while reducing memory usage and improving throughput. Leveraging expertise in computer vision, deep learning, and machine learning, the work was implemented in Python and aimed at enhancing scalability for larger video workloads. The feature was integrated with proper code signing and pull request linkage, laying the groundwork for further optimization and broader deployment without introducing new bugs during the development period.