Early and accurate screening for autism spectrum disorder (ASD) is critical because timely intervention can significantly improve developmental outcomes, yet many individuals remain undiagnosed until much later in life. We are developing an LLM and NLP-based framework that uses narrative generation and transformer-based classification to enable reliable, scalable, and age-appropriate behavioral assessments. The workflow begins with structured survey data, which is transformed into natural caregiver or self-report narratives, validated for consistency, and augmented to capture real-world imperfections. These narratives are then used to train and evaluate transformer-based models, benchmarked against large language model-based classifiers, with performance measured using standard metrics such as accuracy, precision, recall, and F1-score. Together, this framework provides a novel pathway for advancing early and accessible ASD screening through free-text analysis.