A Lightweight Approach: LLM-Driven Classroom Rehearsal on Apple Silicon

This study introduces a software system designed to simulate real-time classroom interactions on portable devices, enabling fully on-device artificial intelligence inference. The objective is to leverage the computational capabilities of Apple silicon devices to seamlessly execute large language model (LLM) and computer vision tasks, supporting teacher training and simulation-based learning scenarios. By integrating live OCR (Optical Character Recognition) processing, CoreML-based YOLO (You Only Look Once) object detection, and LLM inference through the Ollama framework, we developed a platform capable of processing multimodal inputs and delivering instant AI-generated responses in real time. Experimental results confirmed stable performance under concurrent AI workloads, with LLaMA 3.1 (8B parameters) model. Generation speeds scored 17.43 tokens/s with multimodal in-put—slightly lower than the 19.38 tokens/s observed under text-only conditions. Despite the effective performance, limitations remain, including CoreML optimization complexity for LLM models, possible thermal throttle with the peak GPU utilization, LLM size constraints, the need for fine-tuning to enhance simulation realism, and challenges in cross-platform deployment. While improvements are possible, the results highlight the feasibility of deploying advanced AI-based simulation systems entirely on-device, offering a self-contained, low-latency solution for educational applications.