Upserve Case Study

Upserve, originally known as Swipely, is a software and mobile point of sale provider that offers a cloud-based restaurant management platform to restaurant owners across the U.S. The company’s software gives restaurateurs everything they need to know in a single place, with real-time guidance based on sales and trend data. The software integrates with point-of-sale systems and terminals, and it enables restaurateurs to interact with customer spending, social media, and other data.

Upserve captures data streams including credit card payments and menu trends and then turns the data into analytical reports for restaurateurs. “We are devoted to empowering restaurant owners to serve their customers better through data,” says Bright Fulton, director of infrastructure engineering for Upserve. “In addition to providing payment data and reservation system data, we also integrate with online review sites and perform sentiment analysis. We gather all this information and put it into actionable reports and interactive dashboards through mobile apps. We are like a general manager in the cloud for busy restaurateurs.”

Over the past several years, Upserve has strived to provide restaurants with more predictive analysis. “Telling restaurant owners what happened with sales and menu item trends is very important, but telling them what will happen is even more powerful,” says Fulton. “We wanted to discover how we could get prediction capabilities into the hands of our users by taking advantage of machine learning technology.”

As the company considered different machine learning (ML) technologies, it quickly realized that a cloud-based solution would be the best fit. “With the thousands of restaurants we serve, we knew that a machine learning model that works for one might not work in predicting customer behavior in another,” Fulton says. “The idea of creating many custom machine learning models for each customer seemed like a major challenge. We also needed to be able to easily scale the models based on the volume of data coming in. For these reasons, we decided to explore machine learning as a service.”

Upserve decided it wanted to use Amazon Machine Learning (Amazon ML), a cloud-based service that provides visualization tools and wizards to guide developers through the process of creating and training models without needing to learn ML algorithms. “We liked the idea that Amazon ML could enable us to quickly develop ML models on our own,” says Fulton.

Additionally, Upserve was already heavily invested in the Amazon Web Services (AWS) cloud. The organization uses Amazon EC2 Container Service (Amazon ECS) to provision and manage service containers, AWS Data Pipeline and Amazon Elastic Map Reduce (Amazon EMR) for flexible batch processing, and Amazon Simple Storage Service (Amazon S3), Amazon DynamoDB, and Amazon Relational Database Service (Amazon RDS) to store and process hundreds of terabytes of restaurant data. “We have a high level of trust in AWS, especially when it comes to new services,” says Fulton. “We were early adopters of Amazon ECS and Data Pipeline, both of which turned out to be well-placed bets. AWS has always been a great choice for us.”

Upserve began using Amazon ML to create predictive models for its Shift Prep application. Shift Prep integrates table management, point of sale, and other systems to forecast how many people will dine on any given night and which menu items will be popular. Using Amazon ML, Upserve developed more than 100 machine learning models, which collect restaurant data including order information and payment processing data in real time. The models use factors such as the number of reservations scheduled, sales statistics for the same day on the previous year, and customer spend and menu preference histories. Upserve retrains the models weekly.

The company includes the machine learning analysis as part of a daily email sent to restaurant owners through Shift Prep. “Using Amazon Machine Learning, we can predict the total number of customers who will walk through a restaurant’s doors in a night,” says Fulton. “As a result, restaurateurs can better prep and plan their staffing for that night. For example, if more customers are expected, restaurant owners could bring in more employees. In addition, they can use the analysis we provide through Shift Prep to plan specific menu items based on sales history and popularity.

Relying on Amazon ML, Upserve was able to quickly and easily develop and train predictive models. “For us, speed to production was a key factor in choosing Amazon Machine Learning, because we wanted to get predictive analysis to restaurateurs as fast as possible,” says Fulton. “It only took two weeks from the time we decided to use the technology to the moment we started using predictive data in the daily email we send out. And we immediately saw Amazon ML beating the baseline to predicting nightly covers.”

Upserve was able to get up and running on Amazon ML so quickly because of the technology’s ease of use. “The API-centric design of Amazon Machine Learning made it very easy for us to develop and train our models and start getting predictions,” says Fulton. “There wasn’t a lot of configuration required—we used the tools we were already familiar with. Amazon Machine Learning eliminated a lot of development complexity while increasing the accuracy of our predictions.”

The company can also give its customers the ability to increase profitability, because restaurant owners can predict how full their restaurants will be on specific evenings and more efficiently spend labor and food costs. “It’s really important to understand your customer and what they want, and even be ahead of the game and know what they want before they want it,” says Andy Husbands, Chef and Owner of Tremont 647 restaurant in Boston, Massachusetts. “Upserve has really changed how we look at things. It makes it easier for us to really look in the past and see our future, to understand who our guests are and better track them and communicate to them.”

Upserve plans to expand its use of Amazon ML to develop more predictive models. “We expect to grow to thousands of models quickly,” says Fulton. “Eventually, we’d like to expand this to all of our more than 7,000 customers. We are very excited about the future of this technology.”