Worked on the open-edge-platform/edge-ai-libraries repository to enhance DLStreamer with support for Clip-ViT-Base-B16 and B32 models, focusing on robust on-device computer vision workloads. Leveraged C and C++ to refactor VAAPI image processing, improving color space handling and cropping for more reliable inference. Enhanced the CLIP token converter to accommodate varying output blob dimensions, ensuring compatibility with different vision transformer models. Emphasized inference optimization and seamless integration with GStreamer pipelines, resulting in broader model support and improved reliability for edge AI applications. The work addressed both backend efficiency and model compatibility, deepening the library’s capabilities for vision transformer deployments.