Big Data & Analytics Summit - Data Science Challenges @ Lazada
Eugene Yan
A Lazada data scientist shared the real growing pains of scaling a team from 5 to 40 people. Turns out the hardest part isn't the models, it's the humans.
Based on reporting by Eugene Yan — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Eugene Yan spent three years building Lazada's data science function from a scrappy five-person crew into a 40-person team. At a recent summit, instead of pitching flashy use cases, he did something rarer: he talked about what actually broke along the way.
The first fight was over control. Before machine learning showed up, category managers ran things with regex rules and gut calls. When Lazada's ranking algorithms went live, some managers kept manually boosting products they believed in, flagship phones, trending fashion, stuff the algorithm undervalued because of the cold-start problem. Reasonable, at first. But Yan's team ran AB tests and found that once manual boosts crept past a low threshold, well under 100 products, site performance actually dropped. Human judgment helped at the margins and hurt at scale.
The second problem was speed, or rather, the illusion of it. Early on, with eight people and twenty possible projects, the team just picked the ten highest-ROI ones and shipped fast, often with one engineer per project and zero documentation. That worked until it didn't. By year two, unmaintained code and single points of failure, literally one person on vacation being the only one who could fix a production bug, forced a rethink. Yan borrowed Facebook's own lesson here:
My take — AI-written commentary, not fact-checked reporting
"} 오류 detected — need full valid response. AI response cutoff, retry required.
Read more about this at: Eugene Yan