Pre-trained transformers often struggle with multi-lingual code classification due to sequence length constraints and difficulties in explicitly capturing deep structural complexities. To address this for SemEval Task 13, a hybrid neural architecture that fuses CodeBERT’s semantic embeddings is proposed. Handcrafted software engineering metrics is presented, with a Head+Tail truncation strategy to preserve crucial logic in long sequences while simultaneously extracting explicit Abstract Syntax Tree (AST) features via tree-sitter—including maximum depth, branching factor, and cyclomatic complexity. By integrating dense language model representations with explicit structural heuristics, this work provides a robust and scalable solution for enhanced code classification.