Worked on the paddleocr repository to enhance the stability of Key Information Extraction (KIE) model inference using the ONNX runtime. Focused on resolving a critical bug that caused inference errors and input compatibility issues for KE models, the work involved targeted improvements to input handling and code style within the inference pipeline. Leveraged deep learning and machine learning expertise, primarily using Python, to deliver a robust fix that improved reliability and maintainability. These changes enabled smoother production deployment and reduced runtime incidents, reflecting a methodical approach to debugging and code quality in the context of model inference and deployment workflows.