Each year the retail industry collectively loses well over $10 billion because of shoplifting. Now a team of undergrads from the University of Pennsylvania is taking aim at the issue — and doing it in a racially sensitive way. Coming off a recent second-place finish at the Wharton Startup Challenge, Percepta, an artificial intelligence-based shoplifting detector now has the cash backing to launch a pilot product later this summer.
Founded about a year ago by Phillipe Sawaya, Jonathan Mak, and Rahul Shekhar, all of whom will graduate from Penn this spring, Percepta’s algorithms have been trained to spot potential shoplifters based on behaviors while removing all demographic aspects of a person. Neil Gramopadhye, also a Penn student on the Percepta team, calls it a “racially ethical” way to detect shoplifting.
The AI watches CCTV footage that is uploaded from retail stores to the Percepta cloud. It then anonymizes shoppers by removing race, gender, and age. Once the shoppers have been anonymized, the AI observes the movement and patterns of shoppers. It analyzes shoplifting risk and sends an alert to the mobile phone of the in-store operator with the suspicious video for human analyzing.
TAKING RACIAL PROFILING AND BIAS OUT OF SHOPLIFTING
Sawaya, Mak, and Shekhar began working on the software for a senior project. All computer science majors, the trio tapped Alexander Lee — an economics major for business development. Needing a bit more help, Lee called on his fraternity brother and classmate, Gramopadhye in February. The team originally planned on developing the software to look at crimes committed on Penn’s campus to notify the department of public safety. But as they continued their research the trio of Sawaya, Mak, and Shekhar realized their novel algorithm could be applied to many other things, including shoplifting and retail.
“They realized shoplifting was a natural fit for an algorithm that protected some sort of crime committed,” Gramopadhye tells Poets&Quants on a phone call.
“Currently, shoplifting costs retailers about $20 billion each year. Moreover, minorities represent just 20% of shoppers but account for 70% of all shoplifting apprehensions,” Gramopadhye continues. “So we see this disproportionate amount of minorities being apprehended even though they only represent 20% of shoppers. Our goal is to attack shoplifting in a fair and unbiased manner.”
UPDATING AN ARCHAIC SYSTEM
The industry standard now is to put a person in a room for eight hours at a time with up to 50 screens to monitor at once. “When they detect suspicious behavior, they radio people on the floor and ask them to assess the situation and potentially apprehend the suspect,” Gramopadhye says. “But these operators often operate up to 50 screens for eight hours a day. And they end up missing up to 70% of shoplifting instances. And even worse, due to their susceptibility to bias, 40% of the time that they are watching for someone, it’s due to the color of their skin, even if the individual has shown no suspicious behavior.”
Retailers can spend up to $100,000 per store per year on theft prevention personnel, Gramopadhye says. “Our value prop is that we alleviate these issues through our AI-based shoplifting detection software,” he adds.
The Percepta team has trained the algorithms on “tens of thousands of hours of data,” and so far, the AI is producing a 95% accuracy rate with a 5% false positive, Gramopadhye says.
A POTENTIAL PIVOT IF COVID-19 KEEPS RETAIL SHUT DOWN
Despite most retail shops around the country currently being shut down because of social distancing guidelines, Gramopadhye says the team expects retailers to utilize the product as stores begin to open. The team has a minimum viable product they’re polishing to roll out a pilot. “Once stores open up, if anything, shoplifting is going to increase. That’s our thesis,” Gramopadhye believes. “There are a dime a dozen technologies you can adopt to save you money, especially as a retailer. And there are offerings that exist that use an AI face-recognition technology or predictive algorithm that can predict whether a person is shoplifting or committing a crime. But the issue is retailers don’t want to adopt those kinds of solutions because they are afraid of the PR controversy that would arise due to any sort of racially profiling that could happen.”
And if retail is slow to adopt the technology for shoplifting detection, Gramopadhye says the AI is adaptable.
“We can train it on any sort of data to detect anything. The application for this data is more so than just shopping,” he says, adding it could even be used for public health purposes like making sure shoppers are wearing masks and socially distancing.
TAPPING INTO PENN’S ENTREPRENEURSHIP RESOURCES
No doubt the Penn entrepreneurial ecosystem has assisted Gramopadhye and the team. Gramopadhye spent his first year in the research, innovation, and entrepreneurship residential program, which puts like-minded freshmen in the same living hall.
“The great thing about that program is pretty much everyone around you is working on their own startup,” Gramopadhye says. “Penn has so many resources, everyone was doing something. It was this feeling of if not now, then when, you know?”
And now the team has more than $50,000 in cash and services from the Wharton Startup Challenge. Besides earning money by finishing second, Percepta also took home some dough for awards for the best undergraduate team, the new venture collaboration award, and crowd favorite. Gramopadhye is the only member on the team that returns for a final year at Penn before graduating. The team, however, has been accepted into the Pennovation Accelerator which starts this September.