Sporala red del conocimiento
EN
Universidad EAFIT

Maestría en Ciencias de los Datos y Analítica · 2026

Desarrollo de un agente de inteligencia artificial para la gestión preventiva de proyectos en una PMO del sector tecnológico mediante predicción explicable y recuperación aumentada de conocimiento

Project Management Offices (PMOs) accumulate historical records of risks, issues, decisions, and lessons learned that are rarely transformed into actionable knowledge during the planning of new projects. This gap limits the ability to anticipate situations that may affect project success and keeps risk management largely reactive. In response to this problem, this work develops and evaluates a prototype artificial intelligence agent, named the PMO Intelligence Hub, to support preventive project management by integrating explainable risk prediction, retrieval-augmented knowledge, and a conversational interface designed for project managers. The predictive component uses real historical data from a corporate PMO in the technology sector. The model was developed using 576 historical projects, with an overall risk rate of 18.4 %. The pipeline incorporates feature engineering, explicit treatment of missing data as an informative signal, class imbalance handling through SMOTE, isotonic probability calibration, and local explainability through SHAP values. On the test set, which had a positive-class base rate of 18.1 %, the Random Forest model achieved a Recall@0.35 of 0.810, a PR-AUC of 0.326 against a reference value of 0.181, and a Lift@20% of 1.528. The model outperformed the random baseline and obtained better point estimates than logistic regression; however, this comparison should be interpreted in light of the limited size of the test set. Consequently, the results position the model as an early prioritization and decision-support tool rather than as a deterministic classification mechanism. The retrieval-augmented knowledge component indexes 5 329 records of risks, issues, and lessons learned using semantic embeddings. Semantic search obtained better results than a lexical baseline across the three corpus tables, with MRR improvements ranging from +0.20 to +0.35 and ranking quality evaluated through nDCG@5. The generation evaluation using RAGAS achieved an average Faithfulness of 0.86, Answer Relevancy of 0.89, Context Precision of 0.60, and Context Recall of 0.37. These results suggest that the generated responses were mostly faithful and relevant, while also indicating opportunities to improve the selection and coverage of the retrieved context. The architecture is integrated through a conversational agent orchestrated by an LLM (GPT-4.1) with tool-use, capable of retrieving project information, calculating its risk probability, explaining the main contributing factors through SHAP, and automatically triggering an Intelligence Briefing mechanism. This mechanism retrieves similar projects, historical risks, issues, and relevant lessons learned, and presents the results through a functional web interface developed for the prototype. The results support the technical feasibility of integrating explainable predictive models and organizational knowledge retrieval into a unified conversational experience for preventive project planning. The system does not replace the project manager’s judgment; rather, it acts as a decision-support mechanism aimed at prioritizing attention, surfacing actionable historical knowledge, and explaining the main factors associated with the estimated risk. The evaluation was conducted in a prototype environment and did not include production validation, an experimental comparison against an LLM without RAG, or complete manual annotation of the corpus. This work contributes a reproducible architecture for corporate PMOs in the technology sector seeking to evolve from reactive risk management toward a preventive, explainable, and data-driven approach.

Texto completo 68 páginas con texto

Leer la tesis completa Ficha en el repositorio

Palabras clave

Tesauro de su biblioteca