One model, many products
We replaced single-purpose audience classifiers with a transformer-based behavioral model trained on sequences of physical and digital events. It powers daily predictive audiences and provides a shared representation for warehouse enrichment and downstream models for churn, customer value, and next best action. A retailer can buy cheap clicks from people who never visit its stores. We keep bidding separate so campaign incentives do not change what a visit score predicts. In March 2026, Horizon Media reported 48% greater visitation efficiency versus baseline in one pilot combining our predictions with its own audience intelligence and campaign design.
We separate GPU embedding generation from CPU audience scoring. Once the embeddings are ready, daily scoring can run without waiting for GPU capacity. Automatic promotion requires each weekly retrain to clear ranking and lift thresholds across category levels and prediction horizons. Our ML platform tracks which features and datasets each model version depends on.