During November 2025, Itachi contributed to the IBM/vllm repository by developing a Visual Token Masking Performance Enhancement feature. This work introduced caching for visual token IDs and added support for multi-token processing, directly addressing inefficiencies in the model’s handling of visual data. By leveraging Python and PyTorch, Itachi optimized the deep learning pipeline to improve throughput and reduce latency during visual-token-heavy inference. The solution enabled better resource utilization and higher concurrency for model deployments. The engineering approach demonstrated a clear focus on performance optimization, with traceable commits and impact-driven delivery, reflecting a solid understanding of model optimization and machine learning workflows.