Matheus de Medeiros Pereira

At Fácil Espaider I spent two years building multiclass NLP classifiers for legal texts — case updates and procedural publications — using Python, Scikit-learn and TensorFlow/Keras, trained and deployed per client and applied to tens of thousands of records. I designed the pipelines end to end: TF-IDF and n-grams alongside categorical features, Chi-square feature selection, and class-imbalance handling with SMOTE variants.

What sets me apart is the part after training. I owned pre-deployment validation against the model already live (including Top-1 and Top-3 accuracy analysis), acceptance testing, artifact versioning, and post-deployment monitoring. When one release dropped below 1% accuracy in production, I rolled it back the same day and traced the root cause to a preprocessing mismatch between training and serving.

Before that, at the State Comptroller's Office of Santa Catarina (CGE-SC), I automated a financial-reporting ingestion pipeline with integrity checks and email alerting, and built BI dashboards in Power BI, Tableau and OBIEE using Python, SQL, PostgreSQL and Apache Spark. My master's research at UFSC applies deep learning to time-series forecasting — LSTM, TCN, Temporal Fusion Transformer and XGBoost — modelling the link between air pollution and hospital admissions.

I'm based in Florianópolis, Brazil, work in English at C1 level, and I'm available now for remote roles worldwide, including contract work.