Assistant for regulatory compliance and occupational risk prevention through integrated multisource information

Abstract

[Abstract]: This Final Degree Project presents the design and implementation of an intelligent assistant based on locally deployed Large Language Models (LLMs) to support compliance and risk prevention in industrial environments. The system integrates three heterogeneous sources of information: official regulations published in the Boletín Oficial del Estado (BOE), internal organizational documentation (protocols, manuals, policies), and real-time environmental data from Internet of Things (IoT) sensors (Carbon dioxide (CO2), temperature, humidity). To ensure reliable and context-aware answers, the assistant is built upon a Retrieval-Augmented Generation (RAG) architecture. The approach combines semantic search over embeddings stored in Qdrant with a cloud-deployed LLM for answer generation. A React-based web interface provides two complementary views: (i) a conversational chatbot for querying regulations, company policies, and live sensor values in natural language; and (ii) a dashboard for visualizing environmental indicators and historical records stored in PostgreSQL, with data streamed in real time through Kafka. The result is a modular, scalable, and transparent assistant capable of aligning live operational data with regulatory frameworks. The system is especially valuable for technical staff, occupational health and safety officers, and managers seeking actionable insights and regulatory clarity in dynamic industrial contexts.

[Resumo]: Este Traballo de Fin de Grao presenta o deseño e implementación dun asistente intelixente baseado en Large Language Model (LLM)s executados localmente, orientado a facilitar o cumprimento normativo e a prevención de riscos en contornas industriais. O sistema integra información de tres fontes heteroxéneas: regulación oficial publicada no BOE, documentación interna organizativa (protocolos, manuais, políticas) e datos ambientais en tempo real recollidos mediante sensores IoT (por exemplo, niveis de CO2, temperatura, humidade). Para garantir respostas fiables e contextualizadas, emprégase unha arquitectura de RAG. A solución combina busca semántica sobre embeddings almacenados en Qdrant co uso dun LLM exposto na nube para a xeración de respostas. A interface web, desenvolvida en React, ofrece dúas vistas complementarias: (i) un chatbot conversacional para consultar normativa, políticas da empresa e valores de sensores en linguaxe natural; e (ii) un panel de control que visualiza indicadores ambientais e rexistros históricos gardados en PostgreSQL, con datos transmitidos en tempo real a través de Kafka. O resultado é un asistente modular, escalable e transparente, capaz de aliñar datos operativos en tempo real co marco normativo. O sistema resulta especialmente útil para o persoal técnico, responsables de seguridade e saúde laboral, e xestores que precisen obter información práctica e clara en contornos industriais dinámicos.