Cart Intelligence for Food Delivery

Most food delivery platforms run "customers also bought" or collaborative filtering on their cart suggestions. The results aren't surprising: a customer ordering pad thai sees another noodle dish instead of tom yum soup. Someone adding a cappuccino gets recommended a latte rather than a croissant.

Collaborative filtering has structural problems in the food domain. Restaurant-level order data is sparse. A new restaurant doesn't have any co-purchase history. A restaurant that just added seasonal items has no signal on those dishes at all. And even with enough data, co-purchase patterns just reflect what's popular. The most ordered item at a restaurant will get suggested alongside everything, regardless of whether it actually complements the cart.

Cold-start is especially painful. Platforms are constantly onboarding restaurants, and every new one starts with an empty recommendation graph. The first few hundred orders produce suggestions driven by noise rather than any real understanding of what goes with what.

Latimal's suggest endpoint takes a different approach. Send the current cart contents and the restaurant's menu. The API returns items that complement what the customer has already chosen, ranked by pairing relevance.

It works because the model understands food relationships directly. Biryani gets mirchi ka salan instead of another biryani. A burger gets fries and a milkshake. Sushi gets miso soup and edamame. Dosa gets coconut chutney and filter coffee. These pairings hold even on a restaurant's first day on the platform, because the knowledge comes from understanding food directly, without waiting for enough orders to accumulate.

The API also handles subtler pairing logic. If a cart contains only mains, it will prioritize sides, beverages, and desserts to diversify the order. Veg carts get veg suggestions. A cart at a high-end restaurant won't get recommendations from the value tier of the menu. These constraints run automatically without any configuration on your side.

Cart upsell is a volume game. A platform processing millions of orders per month doesn't need a dramatic conversion rate to see meaningful revenue impact. Even a 10-15% improvement in upsell acceptance, which is realistic when suggestions actually make sense, compounds into significant AOV lift across the order base.

The compounding works at two levels. Better suggestions get accepted more often, increasing items per order. And because the suggestions are relevant (sides, drinks, desserts that complement the meal), they're incremental purchases the customer would have considered anyway. Lower friction means higher acceptance and fewer dismissed suggestion trays.

The suggest API handles several dimensions of cart intelligence simultaneously. Category diversification ensures a cart full of mains gets nudged toward sides or beverages instead of yet another entree. Dietary consistency means a fully vegetarian cart never sees a non-veg suggestion, even if the restaurant's bestseller is chicken tikka. Price-tier awareness keeps suggestions within the same range as the rest of the order, so a budget meal doesn't get upsold a premium item that feels out of place.

These constraints make the difference between a suggestion tray that customers engage with and one they dismiss reflexively. When every recommendation feels like something they'd have ordered anyway, acceptance rates climb and the upsell surface becomes a genuine part of the ordering experience rather than an interruption.