Real-Time Personalization Engine for a D2C Brand
The Challenge
The brand's recommendation engine was rule-based: bestsellers by category, manually curated collections, and "customers also bought" associations based on simple co-purchase counts. It ignored real-time behavioral signals — what a user just browsed, how long they spent on a product page, what they added and removed from their cart. The result was recommendations that felt generic and missed obvious revenue opportunities.
Our Approach
Tequity rebuilt the recommendation layer from scratch using a hybrid approach: collaborative filtering for the base model, combined with a real-time behavioral signal layer that updates recommendations session-by-session. User interaction events are streamed through a lightweight event pipeline and fed into a ranking model that re-scores the candidate set in real time.
The serving infrastructure was built for sub-50ms inference at full catalog scale — a requirement for above-the-fold product pages where latency directly impacts conversion. The system supports A/B testing at the algorithm level so the brand can continuously validate improvements.
The Results
Average order value increased 28% in the 90-day period following full deployment. Conversion rate improved 15% compared to the rule-based baseline. Recommendation inference consistently runs under 50ms at peak traffic. The personalization layer now extends across product pages, cart upsells, email recommendations, and post-purchase flows — all served from the same model.
