Artificial Intelligence in Education (AIEd) for Deployment at Scale: Architectures, Evaluation, and Safety

Artificial Intelligence in Education (AIEd) is framed as an engineered system that links formal task definitions, model architectures, safe decision policies, and verifiable operations. This paper introduces a unified task and data ledger covering student modeling, cognitive diagnosis, early warning, assessment, orchestration, and curriculum grounded generation; an architecture selection matrix spanning classical, deep, retrieval augmented, graph, and generative models; a decision blueprint using POMDPs, conservative offline reinforcement learning, and high confidence off policy evaluation; a risk control ledger addressing security, privacy through differential privacy and federated learning, and fairness with calibrated parity; and an evaluation to MLOps playbook that aligns offline metrics and sequential experiments with interoperability contracts, registries, canaries, and rollback. The outcome is a deployment ready methodology that raises mastery gains while bounding latency, cost, carbon, and residual risk across multilingual, multimodal, and low resource settings.