This study explores an artificial intelligence (AI) approach for assessing the geometrical features of flaws in nondestructive testing (NDT) using ultrasonic oscillograms. A validated numerical model, generated through acoustic finite element analysis (FEA), produced ultrasonic signals. The AI model was trained on 525, validated on 113, and tested on 112 oscillograms from models with varied flaw characteristics. Training inputs were parameters derived from ultrasonic signals, and the network's performance was evaluated by comparing its outputs for flaw location, length, and angle with desired values. Statistical analysis, including Root Mean Square Error (RMSE) and Efficiency (E), indicated promising results, suggesting the potential of the proposed AI-based method for estimating flaw geometrical features in ultrasonic NDT.