A compact neural-network model was developed to classify water samples according to quality levels established by CONAMA Resolution 357/2005. The study used a 2018 SEMASA-Itajaí dataset comprising 8,928 hourly measurements of pH, turbidity, apparent colour and local precipitation. After outlier filtering, weighted kNN imputation and min–max scaling, the records were divided into a 70% training set and a 30% validation set. The final classifier, implemented in TensorFlow with two dense hidden layers and trained for 100 epochs with the Adam optimiser, achieved 98.7% accuracy, with misclassifications occurring only between neighbouring regulatory classes. Alongside the modelling, this work also proposes a conceptual design for a future data-visualisation layer intended for deployment in an embedded AIoT monitoring system.