Worked on enhancing the nod-ai/SHARK-Platform by focusing on SDXL inference reliability and performance monitoring. Developed instrumentation to capture timing measurements and average durations for key denoising steps within the UNet component, enabling more accurate latency insights for production environments. Addressed a critical output correctness issue by ensuring data transfer to the host occurs after device synchronization, resolving cases of empty inference results. Leveraged backend development skills with a strong emphasis on logging and performance monitoring, utilizing Python throughout the process. The work improved code traceability and robustness, supporting more reliable and observable SDXL inference in production deployments.