AI-фичи в мобильных приложениях: архитектурный подход — Максим Качинкин, Dodo Engineering

A talk on how to design AI features in a mobile app when the AI ​​lives on the client, not the backend. We'll explore architectural patterns and decision trees: where interaction with the LLM should be located, how to build solutions from one-time requests to the LLM to a custom agent loop (ReAct + tools), and when it makes sense to move to a more mature approach with frameworks like Koog for agent systems and multi-provider systems. We'll discuss the pros and cons of each option, data (RAG vs. tools), production limitations (latency, cost, security, observability), and finally, I'll show you which path we chose in the Drinkit app and why.