The rapid evolution of machine learning frameworks has introduced substantial maintenance challenges for legacy systems. Major transitions across frameworks such as TensorFlow, PyTorch, and OpenCV require large-scale refactoring due to deprecated APIs, execution model changes, and architectural restructuring. Existing automated migration tools primarily address surface-level syntactic updates and fail to account for project-specific abstractions and contextual design patterns. Conversely, naive large language model (LLM) prompting introduces scalability concerns, cost inefficiencies, and risks of ungrounded or inconsistent transformations. This project presents a hybrid migration framework that combines Abstract Syntax Tree (AST) analysis with Retrieval-Augmented Generation (RAG) to enable scalable and context-aware code modernization. The system first performs static analysis using AST parsing to identify deprecated constructs, dependency usage, and structurally transformable components. Deterministic transformations are applied where possible, thereby reducing unnecessary LLM calls and minimizing prediction variance. For semantically complex or architecture-level migrations, the system retrieves relevant modernization patterns from a vector-indexed knowledge base and supplies grounded context to an LLM for structured refactoring. This staged pipeline, AST-driven filtering, retrieval-based grounding, and constrained generative transformation improves efficiency, reduces hallucination risk, and lowers computational cost compared to prompting-only approaches. This work shows that combining static analysis with retrieval-augmented generative models significantly reduces the manual effort required for code migration, though generated code may require some refinement to ensure full functional correctness.

Major Professor: Simanta Mitra