Autism Spectrum Disorder (ASD) is a complex neurodevelopmental condition in which timely and accurate identification is critical for improving long-term outcomes. Traditional diagnostic practices, often based on behavioral observation and caregiver reporting, remain subjective, resource-intensive, and inconsistent across age groups and cultural settings. To address these challenges, this work proposes a hybrid machine learning framework that integrates a Variational Autoencoder (VAE) with an XGBoost classifier for reliable autism trait detection. The VAE component extracts compact latent features from high-dimensional screening responses, while XGBoost utilizes these representations to achieve precise classification. Age-stratified models were trained on datasets covering toddlers, children, adolescents, and adults, ensuring developmental specificity in predictions. Experimental results demonstrate consistently high performance, with accuracy and precision reaching 100% in three cohorts and 99% in the children’s group. Confusion matrix analyses further validate the robustness of the system, showing negligible misclassification rates. In contrast to earlier methods restricted to single data modalities or narrow demographics, the proposed approach achieves scalability, generalizability, and minimal preprocessing requirements. Its lightweight design makes it suitable for deployment in digital health ecosystems, early screening platforms, and adaptive educational environments. These findings highlight the potential of the framework as a clinically reliable and computationally efficient decision-support tool for ASD identification.