My bachelor's thesis: an event-driven multi-agent AI system that analyzes Mongolian Stock Exchange data and delivers personalized investment advice - in Mongolian. Five specialized agents cooperate over Apache Kafka instead of calling each other directly.

Direct agent-to-agent APIs create N×M coupling; routing every message through Kafka brings it down to N+M. Agents fail independently, scale horizontally by adding instances, and any event can be replayed for debugging or model retraining - with Apache Flink handling stateful multi-step execution plans.

The Orchestrator classifies intent with a ReAct loop on Gemini 2.0 Flash and routes work to the Investment, News and Knowledge agents - the latter a pgvector RAG over an MSE company knowledge base. AI queries complete end to end in ~15–20s, API calls in ~200–500ms, and the whole system idles at ~123MB with 99.9% agent uptime.