About
I'm a Technical Lead in ML/AI with deep experience turning state-of-the-art research into systems that hold up under real-world load. My work spans LLM infrastructure, predictive logistics, and the ML platforms that power them.
Currently leading multiple ML/AI teams at MercadoLibre — Latin America's largest e-commerce and fintech platform. I focus equally on rigorous experimentation and the engineering discipline that makes models trustworthy in production.
Pursuing a Licenciatura in Computer Science at UBA alongside industry work. My broader interests include bioinformatics, robotics, and computational neuroscience.
Deep Learning · Generative AI · LLMs · Predictive Modeling · Multi-objective Optimization
Scalable ML Platforms · Online/Batch Prediction · Model Deployment · Observability · ONNX
Python · TensorFlow · PyTorch · XGBoost · LangChain · Hugging Face · OpenAI API
AWS SageMaker · GCP · MLflow · Kedro · TDD · SOLID · Clean Architecture
Experience
Leading multiple ML/AI teams within Latin America's largest e-commerce and fintech platform. Architected the migration of legacy prediction infrastructure, significantly reducing latency and operational costs. Led LLM-based item attribute enrichment using prompt versioning, RAG, and context engineering — with custom evaluation frameworks designed for class imbalance and asymmetric error costs. Built multi-objective shipment time prediction maintaining high on-time delivery rates. Mentored cross-functional teams on engineering best practices, accelerating deployment cadence from monthly to weekly automated releases.
Core predictive modeling for logistics and delivery operations across MercadoLibre's regional marketplace. Developed high-performance models using TensorFlow, XGBoost, and CatBoost, integrated into production via AWS and SageMaker. Designed dynamic batch/online prediction switching strategies to balance cost and performance at scale, enabling millions of daily predictions.
Delivered A/B testing, anomaly detection, and interaction measurement solutions. Built custom data tooling and visualizations to inform business decisions.
Portfolio
01
Predicting airline operational KPIs using least squares regression. An early exploration of applying statistical modeling to real-world logistics data.
→02
A pure Python implementation of the FAdam optimizer to replace standard Adam in microGPT, built entirely without PyTorch. It natively reproduces complex mathematical transforms—including a Fisher Information Matrix (FIM) proxy, adaptive matrix-group clipping, clipped natural gradients, and decoupled weight decay.
→03
Additional work in progress — stay tuned.
Research
Conectia 2025 — Contributor
A custom PyTorch neural network for probabilistic delivery time prediction using beta distribution parameters and the FAdam optimizer (natural-gradient-aware). Compared against NGBoost as a baseline, the approach achieves substantially faster training and inference, a smaller model footprint, and competitive or superior accuracy — with full ONNX export support for production deployment.
Open to collaborations, research conversations, and interesting problems.