A TinyMLOps Framework for Real-world Applications

As devices become smarter, embedding intelligence in microcontrollers and constrained environments is critical. Optimising machine learning models for these tiny devices requires balancing software efficiency, such as accuracy, with hardware constraints like memory and power. We introduce a TinyMLOps-based framework for optimising models across the cloud-to-device continuum. In our approach, cloud resources handle heavy tasks like data labelling and model training, while microcontrollers gather real-time metrics on efficiency and hardware utilisation. Then, some repositories manage models and metadata identified during the optimisation phase, including performance metrics collected directly from the target devices, thus ensuring an accurate exploration of the model space in real-world conditions. Using tools such as MicroPython and MLFlow, our framework enables seamless AI deployment on resource-constrained devices, providing a scalable solution for the future of edge AI.