The creation of Building Information Models (BIM) is driven by cutting-edge software applications, plug-ins, and APIs that constitute the backbone of BIM authoring tools. While free tools and APIs offer visualization and customization options, geometric modelling remains largely restricted to interactive work and proprietary platforms, which sometimes limits flexibility and efficiency. There are still only a few comprehensive workflows that fully automate the reconstruction of building elements from reality-based surveyed data. This paper introduces an innovative reconstruction pipeline developed for the Scan-to-BIM Challenge at the CVPR 2024 Workshop, where it achieved second place in the competition. A deep learning (DL)-driven BIM Module for parametric IFC reconstruction is designed to accurately reconstruct both primary and secondary building elements within a BIM framework, starting from unstructured point cloud data captured via Terrestrial Laser Scanning (TLS). By leveraging DL techniques, particularly Convolutional Neural Networks (CNNs) and Transformers Networks (PTv3), our approach uses late fusion instance segmentation across both 2D and 3D modalities to accurately identify and reconstruct class-specific elements. The pipeline ultimately generates Industry Foundation Classes (IFC) elements, enhancing modelling accuracy, parameter estimation, and consistency in subsequent stages. Results highlight the pipeline’s strong performance on various datasets, underscoring the crucial role of DL in advancing Scan-to-BIM workflows.