Energy-Aware MLOps via Automated Knowledge Distillation: Early Findings and Implications for eXtended Reality

The growing environmental impact of deep learning models has motivated the need for sustainable AI lifecycle management, particularly in production-oriented MLOps pipelines. In this work, we present an early-stage software framework that automatically performs knowledge distillation as a pre-deployment optimization step within an MLOps workflow. The system distills candidate models into more efficient student counterparts and promotes them to production only if they satisfy configurable constraints on predictive performance and energy consumption at inference time. We used the current prototype to perform a benchmark of the currently supported models, including NLP, computer vision, and speech recognition tasks, covering 11 pairs of teacher and student on benchmark datasets such as GLUE, CIFAR, CNN/DailyMail, and Common Voice. Although the number of supported architectures is currently limited, results show a consistent reduction in computational and energy requirements, often accompanied by competitive or improved performance. These findings are particularly relevant for eXtended Reality (XR) applications, where AI models must run on resource-constrained devices such as AR headsets and VR standalone platforms under strict latency, energy, and memory budgets – spanning tasks such as real-time object recognition, speech interaction, and scene understanding. By embedding distillation as a sustainability-aware gate in MLOps pipelines, our framework offers a principled and reproducible path to deploying lightweight models suitable for on-device XR inference. We discuss design choices, current limitations, and the roadmap toward scaling the framework to a broader set of architectures, tasks, and immersive computing scenarios.