Large language models are quickly improving and it has become challenging to differentiate between text written by a human and that by AI and this presents a big worry to academic integrity and content attribution. Conventional stylometric methods tend to miss more complex patterns of semantics, whereas transformer-based detectors do not always take into account stylistic factors that can still be informative to verify the authorship. In this paper, a supervised hybrid authorship verification framework combining a fine-tuned TinyBERT semantic encoder and a stylometric feature extractor that models lexical and structural writing patterns is discussed. In order to enhance the reliability of the classification, lightweight ensemble strategy is used to combine the outputs of both components. The experiments on the official PAN-CLEF 2025 validation dataset have shown that the proposed hybrid model achieves an F1-score of 0.97, ROCAUC of 0.99, c@1 of 0.96, and Brier of 0.07, which is higher than the performance of individual models and classical baseline methods. These findings indicate that semantic and stylistic signals can be used together to supervised AI authorship verification.