At Mercado Libre, I operated more than 50 production credit-risk models, with batch inference covering up to 60 million users. I also enforced release criteria for reproducibility, drift, and model performance.
I led the architecture and data model for an internal explainability platform that reduced the heaviest SHAP run from 5 days to 4 hours. I also GPU-accelerated a SHAP workload from 2 days to minutes, reducing compute cost from about $800 to under $50 per run.
Earlier at Mercado Libre, I proposed and built a Dask-based parallel inference architecture that cut inference time by up to 96%, then packaged it as a model-agnostic library adopted by every production model in the area. At Thomas Processing & Systems, I led the company's first ML initiatives, including object-detection models and a numeric OCR system.
