Fine-Tuned Multi-Model AI Assistance System: A Case Study On LIU

This work presents the design, implementation and evaluation of a fine-tuned multimodal AI assistance system tailored for the Lebanese International University (LIU). The system uses text, audio, image, and video as input, deployed securely on AWS with scalable cloud infrastructure. A dataset of over 7,500 LIU-specific QA pairs was used for fine-tuning this model which achieved 95% training accuracy and 78% validation accuracy through the use of checkpointing. Performance evaluation showed average response times of 1.2s for text, 4–10s for audio/image, and up to 3 minutes for a 30s video clip. Stress testing confirmed 100% availability and an average latency of 0.57s under 200 concurrent users across 40,189 transactions. Security testing demonstrated resilience against Cross-Site Scripting (XSS) and anonymous IP attempts. This work demonstrates that personalized, multimedia-capable AI assistants can be built securely and cost-effectively for educational institutions.