Domain-Specific Fine-Tuning of LLMs for Elevator Emergency Response Systems

Large Language Models (LLMs) have demonstrated remarkable capabilities in natural language understanding and generation. However, their application in safety-critical domains such as elevator emergency response requires domain-specific knowledge and reliable performance, particularly when deployed as offline edge AI systems with constrained computational resources. This study presents fine-tuning of the Gemma-2-9B-Instruct model using Quantized Low-Rank Adaptation (QLoRA) for Turkish elevator emergency scenarios, targeting deployment on embedded touchscreen control panels inside elevator cabins. Unlike cloud-based systems leveraging internet connectivity for retrieval, our edge deployment operates entirely offline, making fine-tuning essential for encoding domain knowledge directly into model parameters. We developed a specialized Turkish dataset containing 1,155 question-answer pairs covering diverse emergency situations. Our evaluation demonstrates significant improvements: ROUGE-1 scores increased from 0.259 to 0.317 (22.49%), BLEU improved by 218.11%, and hallucination rates reduced from 54% to 22% while achieving 49% faster inference. The resulting 4.8GB quantized model runs entirely on embedded hardware without network dependencies, providing immediate, reliable emergency guidance. These results validate parameter-efficient fine-tuning for safety-critical edge AI applications.