Worked on stabilizing and optimizing the PaddleOCR-VL HPS integration within the paddlepaddle/paddleocr repository by addressing a key configuration issue. Focused on routing the vision-language model inference to a dedicated vLLM server, this effort involved patching shell scripts and restoring batch size parameters to improve scalability. Leveraging skills in configuration management, containerization, and DevOps, the changes reduced per-image inference latency and GPU memory usage, aligning deployment behavior with the 1.5 SDK. The work included clear documentation and code review, ensuring maintainability and consistent performance for real-time OCR deployments across different environments, with all updates implemented using Shell scripting.