Three and a half years building data engineering and applied ML systems at Mercado Libre.
Described here only in aggregate, high-level terms — no internal code, architecture, or data is shown.
Queue-driven engine crawling competitor retailers at e-commerce scale.
An orchestration engine that scheduled crawling jobs multiple times a day across dozens of competitor retailers, built on a plugin-based scraping SDK that became the shared foundation for every downstream competitive-intelligence service. Event/queue-driven and horizontally scaled, with years of continuous, incremental production operation.
- 1,145+ production releases of the crawl orchestration engine
- Plugin architecture (search, product, seller, pricing) reused across every retailer integration
- Human-in-the-loop review tooling to keep automated matches accurate at scale
Python
Node.js
Docker
Message Queues
Async I/O
🔒 Built for Mercado Libre — described at a high level only; implementation details are confidential.
Clustering, autoencoders, and classical statistics over 700+ engineered features.
A data-quality-governance pipeline for financial transaction data, combining clustering, an autoencoder, and classical statistical techniques to flag outliers across millions of transactions. Fed real-time alerting and executive BI dashboards, replacing manual review with automated, continuous checks.
- 700+ engineered time-series features driving the anomaly-detection models
- Processed 10M+ transactional records to identify and flag outliers
- Automated schema validation for hundreds of datasets across dozens of business units
Python
Jupyter
Poetry
CI/CD
🔒 Built for Mercado Libre — described at a high level only; implementation details are confidential.
Text classification for miscategorized Open Banking transactions.
An NLP model that classified transactions left in a catch-all 'uncategorized' bucket, using text normalization, TF-IDF vectorization, and class-imbalance handling to make sense of messy, real-world financial text. Iterated from proof-of-concept through to a near-production release.
- Tackled real-world class imbalance in transaction-category data
- Iterated the same model across three stages: research, onboarding, and production hardening
- Directly supported Open Banking / Open Finance compliance reporting
Python
scikit-learn
NLP
Jupyter
🔒 Built for Mercado Libre — described at a high level only; implementation details are confidential.
ETL → training → model-serving, with a sensitive-data governance layer.
A complete MLOps pipeline for a spending-behavior propensity model: data extraction and transformation, training, and model-serving, composed with Hydra configuration management and a dedicated module for handling regulated, sensitive data safely throughout the lifecycle.
- Full lifecycle ownership: ETL, training, and model-serving in one production pipeline
- Dedicated sensitive-data governance module for regulated financial data
- BDD-style system tests validating behavior end to end
Python
scikit-learn
BigQuery
Hydra
pytest (BDD)
🔒 Built for Mercado Libre — described at a high level only; implementation details are confidential.