The master's thesis examines the use of artificial intelligence for optimizing the process of quote preparation in a company operating in the field of geotechnical investigations. The aim of the research was to determine whether large language models, such as GPT, Claude, and QwQ, can effectively support quotation preparers in selecting appropriate items and quantities, thereby reducing time and errors in the process. The thesis includes a theoretical overview of the architecture of modern models and fine-tuning methods. In the practical part, a QwQ 32B model was fine-tuned in the Google Colab environment using domain-specific data and input/output data pairs. The model was tested on 120 inquiries and compared with selected commercial solutions provided by OpenAI and Anthropic. The results showed that locally fine-tuned models do not achieve the accuracy of advanced commercial systems and lag behind in terms of stability and processing speed. The most successful was the Claude Sonnet 4.5 model, which demonstrated a balanced performance in terms of accuracy, interpretability, and response time. The implementation of the selected model into the company’s software solution indicated that the time required for bid preparation can be reduced while simultaneously improving document consistency. The thesis confirms the potential of artificial intelligence for automating repetitive tasks in engineering practice while highlighting limitations related to data protection, the quality of training datasets, and the need for human oversight of the results.
|
