Desenvolvimento de uma plataforma de AutoML

Abstract

In recent years, public and private organisations have intensified their pursuit of data-driven decision-making, yet the construction and upkeep of predictive models remain costly in time, specialised expertise, and infrastructure. Issues such as the derivation and acquisition of new data from existing sources, limited traceability, and shortages of skilled professionals continue to hinder the scalability of the Machine Learning lifecycle. This thesis sets out to design a modular AutoML platform that lowers these barriers by auto- mating the entire workflow, from data ingestion to production monitoring, while avoiding vendor lock-in or the need for advanced domain knowledge. By adopting a vendor-agnostic architecture, the solution seeks to enhance portability, shorten time-to-production, and improve predictive capacity. The proposed architecture begins with an automatic normalisation pipeline that converts structured and unstructured data into a canonical format. It then employs a feature-engineering engine driven by Large Language Models, followed by a multi-objective hyper-parameter optimiser. Probabilistic calibration combined with Conformal Prediction provides uncertainty estimates with formal coverage guarantees. Interpretability layers (SHAP, ICE) expose model rationales, and a self-healing monitoring loop based on shadow models detects drift and triggers retraining without human intervention. Validation in three real-world scenarios: sports sales forecasting, metallurgical copper and zync process optimisation, and academic churn detection, yielded average improvements of +0.13 in R2 and primary-error reductions of roughly 14 %, while cutting delivery times by up to a factor of four relative to manual baselines. These results demonstrate that a unified, LLM-guided AutoML approach enriched with uncertainty management and auto-remediation can democratise reliable and sustainable Machine Learning across diverse domains.