A food delivery app faced a discovery challenge when popularity-based ranking failed to engage users in its curated restaurant widget. The widget, which organized restaurants by tags like Italian or Burgers, showed poor performance compared to other discovery surfaces. Users rarely scrolled past the first 10-12 options, and static popularity sorting couldn't account for contextual relevance or seasonal changes.

The team adapted Uber's Two-Tower Embedding approach, simplifying it for resource-constrained environments. Instead of training from scratch, they leveraged a frozen TinyBERT model for semantic encoding and combined it with restaurant features like price and ratings. User representations were built from recent interactions filtered by the current selection's tag, creating contextual rather than global preferences. Multi-task learning predicted click, add-to-basket, and order behaviors simultaneously.

A/B testing revealed statistically significant conversion rate improvements. The model's flexibility allowed reuse across product surfaces, including ads placement. Key insights included filtering user history by selection tag to capture intent, using negative signals from viewed-but-not-ordered restaurants, and designing for generalization rather than fixed lists. The approach proved that sophisticated recommendation systems can succeed even with limited data and computing resources.