<![CDATA[The rapid development of on-device artificial intelligence (AI) has increased the demand for efficient AI inference on mobile devices, particularly Android devices with limited computational resources, battery capacity, and thermal constraints. Although various optimization techniques, such as quantization, pruning, and adaptive inference, have been proposed, inconsistent evaluation procedures remain a significant challenge to achieving reproducible and comparable measurements of energy efficiency. This study aims to implement a Quality Management System (QMS) framework based on the Plan-Do-Check-Act (PDCA) cycle to systematically optimize and evaluate the energy efficiency of on-device AI inference while maintaining inference performance and model quality. The research employed a design study and experimental approach by deploying lightweight AI models for human presence detection and keyword spotting. The models were optimized using 8-bit post-training quantization, structured channel pruning, and early-exit mechanisms, followed by standardized measurements of energy consumption, inference latency, accuracy, model size, memory usage, and the energy-delay product (EDP) on Android devices. The results indicate that QMS-guided optimization provides a structured framework for improving energy efficiency by reducing computational workload and memory traffic while preserving inference reliability. In particular, post-training quantization combined with hardware acceleration reduced energy consumption, whereas adaptive inference mechanisms improved responsiveness under varying battery and thermal conditions. In conclusion, integrating QMS principles with on-device AI optimization establishes a reproducible, traceable, and continually improvable framework for developing energy-efficient AI applications on Android devices.]]>
