Authors - An Doan Van, Dong Nguyen Doan, Quynh Tran Duc, Thuan Nguyen Quang, Bao Phan Gia, Hieu Doan Minh, Van Khanh Doan Abstract - Performance bottlenecks in Python programs arise from a wide variety of sources, and no single technique reliably catches them all. This paper proposes CodeForge, a sequential three-stage optimization system that unites deterministic Abstract Syntax Tree (AST) inspection, CodeBERT embedding-based retrieval, and Gemini LLM-driven rewriting into one end-to-end pipeline. A rule engine in the first stage pinpoints well-known structural problems; a neural similarity search in the second stage captures harder-to-spot variants; and a Gemini LLM in the third stage performs the actual rewrite, guided by a structured hint block assembled from both preceding stages. Before any result is returned, a configurable validator rejects changes that fail minimum speedup, memory, or complexity criteria. Alongside each accepted optimization, a composite confidence score and a plain-language rationale are produced. Tests on six representative Python patterns show that hint-guided LLM prompting raises successful detection from four to six out of six cases compared with unguided prompting, while the validation layer blocks every harmful transformation in the test suite. The system is available as a FastAPI REST service accepting both raw source text and uploaded .py files.
Open Zoom