“Lord of the recommendation models, Two Tower for All(egro)”
Building effective recommendation systems for a massive e-commerce platform like Allegro.com—with over 20 million active buyers and a highly dynamic catalog —presents formidable challenges. How do you deliver diverse, high-relevance suggestions across dozens of placements while keeping maintenance low? We present a unified, content-based recommendation system built on the Two Tower retrieval framework. We demonstrate how this single architecture is adapted to serve three distinct user intents: similarity search, complementary suggestions, and inspirational content discovery. Our solution handles dozens of thousands requests per second at production scale. Insights from two years of continuous A/B testing confirm statistically significant gains in business metrics, proving that flexibility and scalability are key for product development at Allegro scale.
Bio
Paweł is a Machine Learning Engineer at Allegro, currently working within the Recommendations team. His focus is on inspirational recommendations—helping users discover products they didn’t know they needed by broadening their horizons—and experimenting with agentic approaches to enhance ML model capabilities. Previously, as part of the Learning To Rank team, he was responsible for product ranking in the search engine, where his optimization efforts successfully reduced the total mileage on users’ mouse scroll wheels. In his spare time, he trades data peaks for real ones, hiking in the mountains or traveling around the world.