Maple, a David Chang-backed restaurant in New York City, doesn’t have any tables, cash registers, or waiters. Instead, its customers order meals through its website or mobile app, and a fleet of bike couriers deliver them. By eliminating the dining room and bringing meals to you, Maple is betting that it can sell more meals per hour, using less real estate, than a traditional restaurant.
The current gold standard for zipping patrons through a lunch line (what the industry calls “throughput”) is Chipotle. According to its 2014 annual report, Chipotle manages to serve 300 meals per hour—a transaction every 12 seconds—at its best-performing locations, and the chain is so obsessed with its productivity that it assigns employees efficiency roles with names like “linebacker.”
When Maple launched its first location in April, it served around 50 meals per hour at peak times. Less than a year later, on average it is now serving 800 meals per hour from each of its four kitchens. A few days before I visited in February, it had set a new record: 1,100 meals cooked and delivered in one hour.
Some of Maple’s insane improvement in meal-per-hour productivity can be chalked up to increasing demand—more people know about Maple now than did during the first week after it launched—but the company has also invested heavily in technology in hopes of beating the efficiency of brick-and-mortar restaurants like Chipotle. While most food-delivery companies use smartphones to connect customers with couriers, Maple owns the entire restaurant and delivery system, which means it can also use mobile tech and data science to optimize its entire workflow.
Cooking, assembling, and dispatching meals on demand, at scale, is a complicated logistics puzzle. More specifically, it is a “stochastic open capacitated vehicle routing problem with soft time windows and multiple trips,” according to Maple’s CTO, Dan Cowgill. In layman’s terms, that means that Maple faces a math problem of plotting delivery routes efficiently, like a delivery company, combined with the problem of only being able to carry so many meals per trip, combined with the problem of a customer not showing up to receive a delivered meal every once in a while. All of which is made more complicated by a perishable product prepared on demand, and the potential that delivery people could return to a different Maple kitchen than the one at which they picked up their last order.
When Maple started, it based its kitchen technology on existing restaurant procedure. Its kitchen operating system centered on an app with a drag-and-drop interface that moved orders from “cooking,” “plating,” and “ready to be bundled for delivery” status, much like a paper ticket that moves down a cook line at a diner. But, says Maple cofounder and COO Akshay Navle, “It was a total waste of time to have the cook do anything other than cooking.” Now, there’s a separate app that shows cooks the predicted demand for orders (using simple machine learning techniques that base forecasts on past performance and menu mix), another that helps workers who plate dishes keep the system updated on the number of completed meals available, and a third that shows workers who bundle orders for delivery what to pack in which courier’s bags. Nobody needs to drag and drop.
Within this app ecosystem, the “bundling algorithm” is the trickiest to get right. Dividing delivery orders between couriers involves plotting out their optimal route and balancing factors like wait times at each building. Orders going to the same building should obviously be packed together, but how long can a courier wait at the kitchen for another order at the same address before missing the time at which Maple had promised to deliver? Which buildings should be on the same route?
Originally, Navle spent rush hours grouping Maple orders for delivery at his computer monitor. In mid-February, Maple switched to an automatic system that Cowgill built. It takes into account how many riders are in each kitchen, how fast cooks are at making each menu item, what demand looks like historically for a day like today, thousands of promised ETAs, the average to and return time for the last 100 runs to a particular building, and anything building-specific (freight elevators, doorman, etc.) that might affect delivery times. Now, not only is Maple becoming more efficient with increasing demand, which means it can more often send multiple deliveries to the same addresses at once, but it can be smarter about the routes its couriers take.
Maple’s kitchens, around 3,000 square feet each, aren’t much bigger than Chipotle’s stores, which average 2,500 square feet. They’re each staffed with about 20 people, while Chipotle stores on average run with a crew of 24, plus managers. With no need for foot traffic, Maple avoids paying a premium for prime retail locations. But Maple hires an extra 50 bike couriers per kitchen—a significant expense.
That expense, however, makes more sense if each Maple kitchen can cover a much larger area than a Chipotle location. “Chipotle has something like 40 locations below Central Park,” says Maple CEO Caleb Merkl. “We’ll be able to service that same meal volume by only needing to build five to seven locations.” But the area each Maple kitchen can cover is limited by how many meals it can produce and deliver per hour.
That’s why continuing to invest in technology that can squeeze every last bit of efficiency out of its modest space and crew is important. Maple recently introduced, for instance, ordering windows that serve a dual purpose of allowing customers to schedule their lunch delivery at certain times and, at some point, helping to manage demand. The app could, for instance, allow people in a building where deliveries are already scheduled to order lunch, while telling a customer in an address not already scheduled in a route that lunch is sold out. Or, it could accept your order for instant delivery of a salad, which can be prepared in minutes and can join a trip to your address that’s almost ready to leave the kitchen, but tell your coworker who wants to order a chicken breast, which takes longer to cook, that his order will need to be slotted into the next hour’s trip.
With data science and smartphones, possibilities for increasing efficiency seem endless. As it scales, Maple even plans to coordinate its couriers so that they don’t need to come back to one kitchen to pick up orders—they can return to a different, closer hub, or receive a new order from another courier they’ll pass on the way.
As Navle says, “We’ve built a machine.”