Cyberbullying has become a common issue in the computer age and has affected people of all ages, leading to serious emotional and psychological damage. As social media continues to grow at a high rate, it is important to identify and label abusive materials at a very early stage. The proposal suggests a scalable AI-based framework for multi-class cyberbullying detection based on transformer-based models, BERT and RoBERTa. This system is trained on a subset of curated Cyberbullying Classification data on Kaggle, comprising about 40,000 labelled tweets across five categories: age, religion, ethnicity, gender, and non-cyberbullying. The approaches include text-intensive preprocessing, label encoding and cleaning, and weighted focal loss alleviates class imbalance, and RoBERTa can freeze its layers to efficiently fine-tune them. Model performance can be assessed using precision-recall curves, which take into account accuracy, recall, F1-score, ROC-AUC, and average precision (AP). In comparison to RoBERTa's 94.71%, BERT's accuracy, precision, recall, and F1-score are 94.97%. In addition, it can be understood with the aid of LIME (Local Interpretable Model-Agnostic Explanations), which highlights the key terms that contribute to predictions and enhances transparency. To identify and categorise cyberbullying early and provide practical insights for creating safer online environments, the results validate the premise that transformer-based models effectively represent contextual and semantic patterns.