Worked on enhancing input data handling and preprocessing reliability in computer vision pipelines, focusing on the ultralytics/ultralytics and alibaba/MNN repositories. Addressed a critical bug in OpenVINO preprocessing by ensuring correct tensor data assignment and robust memory copying from resized frames, which reduced inference errors and improved production stability. In the multiPose demo for alibaba/MNN, implemented input tensor dimension type detection to support both CAFFE and TENSORFLOW formats, updating tensor resizing logic to accommodate multiple input types. Leveraged C++, Caffe, and TensorFlow to broaden input compatibility, streamline integration for downstream models, and maintain alignment with repository standards throughout development.