AI Recommendation Engine for Romanian E-Commerce

RetailMax SRL was a growing Romanian e-commerce player—good product, good logistics, but commodity-level merchandising. With 50,000+ products and no personalization, customers spent 15+ minutes browsing before purchasing. Cart abandonment hit 45%.

Building the Recommendation System

We built a three-pronged recommendation engine: collaborative filtering to find similar users, content-based filtering to match product attributes, and contextual signals like browsing patterns and time-of-day behavior. The system scored candidates from all three approaches and ranked them together.

The architecture was deliberately simple: recommendations were computed in real-time using Redis caching and a lightweight ranking service. No 12-hour batch jobs. No complex feature engineering initially. Just solid fundamentals that scaled cleanly.

Deployment Strategy

Rather than a big bang launch, we rolled out incrementally. First the sidebar on product pages, then the homepage, then the checkout flow. Each placement had its own A/B test, which let us understand exactly where recommendations created value.

Within 4 weeks, we saw measurable lift: 58% increase in recommendation click-through, 19% increase in average order value on orders that included a recommended product. The team at RetailMax took over operations and continued optimizing.

Lessons Learned

Simple systems that work beat complex systems that fit the theory better. We debated fancy neural collaborative filtering approaches initially. In the end, a well-tuned linear combination of simpler signals proved more reliable and easier to maintain. Start there.

Also: data wins. The quality of user behavior data, product metadata, and feedback loops determined the ceiling on recommendation quality far more than algorithm choice.