Model-Driven Engineering (MDE) fosters the active use of models throughout the software construction process. Models are typically built using domain-specific languages, which are themselves described via a meta-model. While MDE yields advantages due to its possibilities for automation, modelling and meta-modelling remain complex tasks, which would benefit from intelligent assistants.

In this paper, we propose an assistive approach based on Large Language Models (LLMs) to help in the construction of both meta-models and domain-specific models. The approach works at the abstract syntax level, hence being independent of the (graphical, textual) concrete syntax of the modelling language. We have realised the approach as an intelligent assistant for EMF, called EMF-Kaizen. The assistant suggests model fragments that satisfy the user requests and can be incorporated to the model under construction (e.g., in the default tree-based modelling editor) via drag&drop. It also maintains and persists the modelling assistance chat sessions, together with the suggested model fragments. We report on an evaluation of EMF-Kaizen over five meta-models (including Ecore) at different modelling stages and usage scenarios, showing its usefulness.