From Dialogue to Design: GenAI-Based Automation of Parametric Modeling and Sizing Tasks in CAD Workflows

• Fine-tuned LLMs enable parameter extraction, sizing, and FreeCAD macro code generation. • Small datasets (approximately 2,000 samples) are sufficient for high-accuracy design automation. • Fine-tuning with extended hyperparameters improves quality without overhead. • Early stopping reduces training time while preserving model performance. This study explores the automation of engineering design tasks using Generative Artificial Intelligence, focusing on the dimensioning machine elements and generation of their respective CAD models, specifically bolted connections, in the open-source software FreeCAD. Three system architectures were developed and evaluated: All-in-One, Chatbot-Designbot, and Chatbot-Calculator-Code Generator. These frameworks integrate Large Language Models, such as GPT-2 and CodeGen, which were fine-tuned using Parameter-Efficient Fine-Tuning and Low-Rank Adaptation. While the All-in-One architecture consolidates all tasks into a single model, the Chatbot-Designbot and Chatbot-Calculator-Code Generator architectures decompose the process into specialized modules for dialogue interaction, parameter extraction, part dimensioning, and CAD code generation. The evaluation results show that the Chatbot-Calculator-Code Generator configuration achieves the lowest overall error rate for the four steps combined (2%) with a total training time of 63.4 minutes. This configuration outperforms the Chatbot-Designbot (3.96%, 117.7 minutes) and All-in-One (53%, 24.2 minutes) architectures. These findings demonstrate that compact, fine-tuned Large Language Models can enable accurate and efficient design automation, even with limited data. This work establishes a methodological foundation for scalable, Generative Artificial Intelligence -driven CAD systems and interactive engineering design workflows.