Two years ago, software engineer Georges Duverger started writing down everything he ate.
He had spent the last several years working in high-stress startups and eating poorly, and he said as he got closer to 30 his body wasn't reacting to the food the way it used to.
"At that point in time, I had started to gain a little weight and wanted to get it under control," he said. "Something like two thirds of Americans want to lose weight, I think it's a common thing to want."
He began with calorie counting apps like MyFitnessPal and Lose It!, but could never find one he could stick with for more than a few days. With the rise of automatic fitness trackers like Fitbit, Duverger wanted to build a way to track nutrition that was just as seamless.
He developed his own system, initially emailing himself lists of the foods he ate each day, and later settling on the most straightforward mechanism he could think of: a simple text.
More than 3,000 entries and 230 quantified espressos later, Duverger lost the weight—but the experience made him realize how difficult is it to get nutritional insight beyond just calories. He began to develop Fitmeal, an on-demand nutrition service that replaces calorie counting with a computer program that teaches itself about the user's diet over time through machine learning,
The app allows users to text anything and everything they eat to a phone number that automatically replies with nutrition facts, including calories, fats, carbohydrates, and protein. He culls the nutritional info from a variety of online sources, including a USDA database.
The longer users track their foods, and the more information it has to analyze, the more accurate it should be
The machine learning aspect allows users to be as specific or as vague as they want, Duverger said. For example, a user might want to text an exact amount of food, like, "12 ounces of cheese" to Fitmeal, but he developed the program in a way that the user could also type "three slices of cheese" and the app will still understand.
"A machine learning algorithm is used to make that abstraction," he said. "It looks at a lot of meals online, tries to figure out all the meals that have cheese in them, and then how much cheese is usually meant by a 'slice.'"
Duverger said he was first exposed to machine learning at Hunch, a company later purchased by eBay that creates recommendations for users based on their interests. He said he has built Fitmeal with some preexisting core tools like natural language processing, but has not found many machine learning programs in the fitness realm.
"I'm not really reinventing the wheel in terms of how these things are done, but I'm applying these tools to the field of nutrition, and I haven't found any tools in libraries online to do that, so I had to create all of that from scratch," he said.
For now, Fitmeal just sends users immediate information about the nutritional value of each food or meal texted to it, but Duverger said he is experimenting with more comprehensive feedback, like daily emails chronicling categories of foods and weight loss over longer periods of time. He said with machine learning, the longer users track their foods, and the more information it has to analyze, the more accurate it will be.
The accuracy levels vary. A bowl of oatmeal I texted the app the other day, for example, was 50 calories off of the actual calorie count. However, the information can occasionally be a bit off. For example, I texted my lunch, turkey and cheese, to the app and it told me it was 223 calories, when the actual calorie count was 163, based on the labels of the food. I texted it again with more specific info to see if it would improve, but it sent me the same calorie count and then a third text adjusting the number to be even higher than the initial, inaccurate count.
Duverger said the incident is a good example of the issue of overestimating and underestimating he is trying to refine in the system. He said in the future he may experiment with user feedback, like "too high" or "too low" to help the program better learn averages. He said he is trying to balance interaction with simplicity.
"I don't want that to take away from the user feeling that it's easy, I don't want to make it as difficult as the other apps, but it's definitely something I want to explore," he said.
The app is still in beta, and has only about 70 users at the moment, but prospective users can sign up for a waiting list on the service's website. According to a 2013 Pew Research Center survey, 60 percent of adults already track their weight, diet, or exercise routine, which means there is a large potential market. However, there are already a number of similar apps out there, such as MyFitnessPal, Lose It!, and My Diet Diary. Duverger said he isn't worried about entering a crowded market, however.
"The space is big," he said. "Two thirds of American adults are overweight, so there is a lot of room for everybody. I don't think it's a matter of one is going to take over another."