Deepcall

Deepcall is a meal-recommendation engine built for one sharp business problem: keeping customers from quietly churning. In food ordering, a lapsing user rarely announces they are leaving — they simply stop coming back. Deepcall predicts the next basket a user is most likely to order and uses that prediction to surface a targeted offer for their exact dish, reaching at-risk and recently-lapsed customers before their loyalty is gone. The name is the mandate: a deliberate call back to the users worth re-engaging.

Ordering history is modeled as a graph. Each user is a root node whose interactions branch outward — user → basket → menu item — so a single customer’s history becomes a connected subgraph rather than a flat event log. As new dishes and orders arrive, the graph grows fresh item nodes and edges in place, keeping the menu and customer base a living structure rather than a fixed vocabulary.

On top of this sits a custom SASRec self-attentive sequential model over learned embeddings, predicting the next order with a structural bias toward well-formed baskets — building, main, sides, dessert — rather than isolated items. Interpretability is not bolted on after the fact: linear probes at every item prediction make each recommendation legible, so a candidate dish carries a readable reason instead of an opaque score. The system trains continuously, folding in new orders as data arrives so it tracks shifting tastes instead of decaying against a stale snapshot.

Trained on six years of ordering data — 14 million orders across 4 million users — Deepcall delivers a 115% lift over baseline, an NDCG@10 of 0.75, and a 15% reduction in churn.

Built as lead developer. The repository is private — the work is proprietary.