Constrained Hybrid AI for Smart Classroom Audiovisual Diagnosis

Smart classrooms are consumer-scale cyber-physical systems in which heterogeneous audiovisual (AV) failures must be diagnosed and resolved within a strict pre-class window. Conventional rule-based self-checking is fast but brittle under closed-world assumptions, while purely data-driven AI is difficult to deploy safely due to limited training data and the risk of unsafe actions on physical devices. This article presents a constrained hybrid diagnostic architecture, studied as an early deployment and integration effort, that combines deterministic rule-based self-checking, non-intrusive physical verification, and a large language model (LLM) used only as a constrained diagnostic planner. The system separates diagnostic reasoning from physical execution: rules and LLM proposals are represented as auditable check actions, all actions are routed through a policy enforcement point, and unsafe or high-impact requests are rejected before reaching devices. Controlled fault-injection experiments on a representative smart classroom testbed covered ten predefined fault categories and 100 repeated runs, while the architecture was deployed in 70 real classrooms. In this controlled setting, the constrained LLM planner achieved 99% accuracy of root-cause identification, generated safe diagnostic paths when rule priors were incomplete, reduced a representative symptom-driven diagnosis from 38 static checks to 7 checks, and produced 14 unsafe or out-of-scope requests, all of which were successfully blocked by the policy layer.