The proliferation of IoT devices has significantly increased global energy consumption and carbon footprint due to the reliance on computationally intensive Machine Learning (ML) techniques. Traditionally implemented on high-end embedded devices or exploiting cloud computing, these applications further exacerbate environmental impacts. As IoT technology advances, there’s a shift towards smaller, low-power devices, necessitating decentralized computation. This shift has given rise to Tiny Machine Learning (TinyML), enabling ML tasks on edge devices. However, to fully exploit TinyML, integrated frameworks are essential for efficiently developing and deploying smart applications, especially on devices operating intermittently on harvested energy. The Green Machine Learning for the IoT (GEMINI) project aims to address these challenges by creating a framework for sustainable TinyML applications on battery-powered and batteryless energy-harvesting edge devices. GEMINI targets shifting computational workloads from the cloud to the edge, optimizing latency, bandwidth, energy use, and privacy. By integrating energy harvesting methods, zero-power communication protocols, and intermittent computing strategies, GEMINI seeks to ensure robust ML inference under intermittent power conditions. The project aims to include efficient data collection, ML model generation, and online learning directly on edge devices, reducing reliance on constant cloud connectivity. The open-source framework intends to offer a comprehensive toolkit for developing sustainable IoT solutions.