Ecosistema multi-agente basado en LLMs para atención estudiantil integrada en servicios universitarios

Abstract

Abstract: This thesis presents and evaluates SUEMas, an LLM-powered multi-agent ecosystem designed to integrate and enhance university student support services. It tackles service fragmentation and manual workload through a hybrid architecture that combines a central orchestrator (LangGraph) with domain-specialized agents (certificates, job vacancies, résumé generation, knowledge, buses, scholarships). The system emphasizes governance and operational safety via a closed-world approach, deny-by-default policies enforced by a Dynamic Tool Registry, and conversational continuity using session caching (TTL/SWR). To mitigate factual hallucinations in informational flows, SUEMas incorporates hybrid RAG (semantic + keyword retrieval) and preserves source traceability. Validation includes deterministic testing of policies and cache behavior, RAG evaluation (retrieval and grounded generation), and multi-turn conversational coherence, yielding satisfactory results and response times suitable for a chat-based user experience.