How an AI assistant evolved into a production system: from NLU to orchestration and reliable message processing

This article provides a detailed analysis of how an AI assistant evolved into a full-scale production system. The author examines key stages of architectural transformation, including dialogue state management, separation of business logic and communication channels, and the integration of NLU and LLMs. Special attention is given to orchestration, interaction with human support, and the implementation of Inbox/Outbox patterns to ensure reliable message processing. The article demonstrates how creating a unified lifecycle for user requests becomes a critical aspect, where architectural decisions often prove more significant than the AI algorithms themselves. This material is useful for developers and architects working on scalable conversational systems.