Adapting Large Language Models for Code Generation in Proprietary Contexts: An Industrial Case Study

Abstract

This dissertation investigates the adaptation of large language models for proprietary code generation within industrial contexts, with a case study conducted in collaboration with Critical Manufacturing. The work addresses the limitations of generic large language models when applied to private codebases, including the lack of alignment with proprietary application programming interfaces, coding conventions, and compliance requirements. It examines strategies for improving domain alignment without full-scale fine-tuning, including retrieval-augmented generation, static context integration, and iterative self-correction mechanisms. It proposes and implements an iterative pipeline for code generation that progressively integrates retrieval-augmented generation, static context, system prompt refinement, and compilation-based feedback output refinement. A reproducible benchmark of twenty problems was developed to evaluate these strategies, combining metrics of syntactic validity, functional correctness, and developer effort. Developer effort was quantified through the Manual Coding Ratio, defined as the proportion of the final solution that must be manually authored after generation. The evaluation showed that compilation success increased from 25\% at the first attempt to 45\% after three attempts, while execution success rose from 5\% to 15\%. Corrective effort, measured by the Manual Coding Ratio, decreased from 45.78\% to 41\%. Although absolute execution success remained limited, these results confirm that retrieval and iterative refinement improve the practical usability of large language model outputs in proprietary environments. The study contributes a structured methodology for assessing large language model-assisted code generation in industrial contexts and provides evidence of the trade-offs between cost-efficiency, performance, and compliance when deploying large language models under resource and privacy constraints.