Recommendation systems reportedly drive 35% of Amazon's revenue.

Netflix says 80% of watched content comes from recommendations.

Yet most ML engineers have never built one from scratch.

The architecture has evolved dramatically. The modern approach uses a two-tower model: one tower embeds users, another embeds items. A retrieval stage quickly narrows millions of candidates to hundreds. A ranking stage carefully orders those hundreds. Then business rules apply on top.

What's new in 2026: transformer-based sequential recommendation (modeling the sequence of your interactions, not just individual preferences), multi-objective optimization (balance engagement, revenue, AND user satisfaction), and LLM-powered recommendation where language models reason about why you might like something.

The fascinating challenge is that recommendations face a tension most ML systems don't: the model's predictions change user behavior, which changes the data, which changes the model. It's a feedback loop that can spiral into filter bubbles if you're not careful.

If you want to understand the intersection of ML, product thinking, and business impact — build a recommendation system. There's no better training ground.