This paper examines the evolution of the neural network module architecture of the Mirera Digital Educational Platform in connection with the expanding range of applied artificial intelligence models. It is shown that the original LLM-oriented orchestrator core did not allow models of other modalities to be integrated without substantial reworking. A universal multimodal architecture is described, including a service manager abstraction based on a unified contract, configurable request-processing chains, and a unified stream of output artifacts. Examples of integrations are presented. It is shown that the proposed architecture enables the integration of new model types without modifying the router core and provides a basis for the further development of multimodal educational scenarios.