Intelligent AI-based Framework for Comprehensive Risk Identification and Management in Construction Projects

This research proposes an AI-based architecture to detect risks and manage feasibility of construction or infrastructure project designs. It consists of three main components: (i) PDF interpretation for risk parameters with CLAHE domain-specific image preprocessing technique and OCR-based risk detection model, (ii) hybrid interpretation with two neural networks: embedding layer with pre-trained BERT language model and classification layer with LightGBM model, and (iii) feasibility analysis model for building floor areas estimation and building type selection for soil and environmental conditions. Four innovations were implemented: domain-specific CLAHE preprocessing, dual-module riskfeasibility ensemble with probabilistic conflict handling, optimized FP16 embeddings on CUDA, and UniStrided-1 processing. Experiments with 12,500 engineering reports show 88.9% risk detection accuracy, 75.5% MOTA for risk tracking, and 86.8% mIoU for feasibility segmentation. The framework analyzes 19.5 reports per minute on a Tesla T4 GPU and reduces manual analysis time by 68%.