This thesis develops an integrated intelligent framework for ARBoard, an Augmented Reality (AR) assistance platform designed to optimize manual operations and diagnostic workflows on Printed Circuit Boards (PCBs). While the broader system includes spatial mapping and visualization, this work contributes a software backend that unifies a specialized text-and-voice AI assistant with automated board data extraction. These two functional layers are integrated via a centralized relational database. The Reasoning Layer introduces a custom Benchmark Engine designed to quantify tool-calling performance across various open-source architectures, including the GPT-OSS, Qwen, and MiniMax series, to compare and evaluate the models. To bridge the domain gap in technical PCB documentation, a Teacher-Student distillation pipeline was implemented, using high-reasoning models, to synthesize high-fidelity, bilingual synthetic dataset.