A Serverless Cloud-to-Thing Framework for TinyMLOps Workflows

The deployment of Machine Learning (ML) across the Cloud-to-Thing continuum is a significant challenge, particularly when considering the heterogeneity of the available devices. This work introduces a general- purpose framework for Tiny Machine Learning Operations (TinyMLOps) that enables the orchestration of ML workflows in distributed, serverless environments spanning cloud, edge, and Internet of Things (IoT) nodes. The architecture follows an event-driven model in which each node—depending on its capabilities—implements a minimal mandatory set of components and an optional set of extended functionalities. Nodes advertise their capabilities and collaboratively fulfill MLOps tasks by handling requests they can satisfy. This decentralized approach allows for dynamic, context-aware distribution of operations across networks of heterogeneous resource- constrained devices. The framework is validated by means of a prototype implementation and early experiments involving STM32-based microcontrollers and Raspberry Pi edge devices.