Developed and integrated advanced multimodal capabilities across jeejeelee/vllm and huggingface/transformers, focusing on the IBM Granite 4.1 Vision model. Work included implementing end-to-end image and text processing within vllm, addressing decoding bottlenecks by optimizing buffer management, and ensuring production readiness for scalable multimodal workflows. In huggingface/transformers, delivered a modular, built-in Granite4Vision model with automated registration, comprehensive tests, and robust documentation. Enhanced image feature extraction and improved configuration and weight-loading reliability for downstream usage. Leveraged Python, deep learning, and computer vision expertise to deliver stable, maintainable code that supports both research and production deployment of multimodal models.