Javier Mansilla has been building machine learning systems since 2010 — before neural networks became the industry’s default hammer. As co-founder of Machinalis, later acquired by Mercado Libre, he watched a small team evolve into the ML backbone for a platform serving nearly 100 million users across Latin America. Now directing ML, AB testing, and user tracking for the NASDAQ-listed giant, Mansilla has a rare vantage point on what actually works at scale.
His assessment is refreshingly blunt. ML’s biggest wins in e-commerce aren’t the flashy ones. Fraud prevention, he says, is where machine learning “made the impossible possible” — a domain where rule-based systems simply couldn’t keep up with adversarial actors. The second tier involves UX improvements like discovery and serendipity in product listings, where ML delivers a level of personalization that would be prohibitively expensive to build manually. Search, recommendations, ads, credit scoring, moderation, forecasting, logistics — every core unit at Mercado Libre now runs with ML optimizing at least one fundamental metric. Even infrastructure reservation gets the ML treatment.
The challenges, though, are structural. E-commerce ML is inherently probabilistic, and Mansilla is candid about the tension: when you’re scaling market share year over year, a 95% accurate model still means real users hitting real failures. The long tail matters. Adding complexity is the coordination problem — managing data flows across marketplace, logistics, credit, insurance, and brick-and-mortar payments simultaneously. That’s a governance nightmare most companies never solve.
His advice for teams integrating ML cuts against the hype cycle. Start with the simplest baseline. Measure with/without impact explicitly. Know the boundaries of expected gain before investing. And stop ignoring technical debt in pursuit of short-term wins. That perspective — earned from 14 years in the trenches — echoes what pragmatic ML leaders have been saying for years, but it lands differently coming from someone operating at Mercado Libre’s scale. The industry, he notes, is finally solving the “slow, repetitive, boring pieces” of the ML pipeline, which means shorter model-building cycles and faster time to market. That’s where the real leverage is.
