At Cadastra, working with Grupo Boticário, I designed end-to-end ML pipelines on Google Cloud Platform that automate retraining for nine models. I also architected an event-driven pipeline for on-demand retraining and prediction.
I integrated GitHub Actions CI/CD for model deployment and refactored legacy time-series forecasting models to improve maintainability and inference speed.
At BWTech, I productized and monitored on-premises machine learning systems, developed Airflow pipelines, and improved forecasting, clustering, and anomaly detection models. I also developed components of a RAG solution for telecom documentation retrieval using LangChain and Pinecone.
My master's research at Universidade Federal de Minas Gerais examined bias in AI toxicity models, including their reliance on isolated keywords rather than context. My work and research span NLP, fairness, and production machine learning systems.
