Why FitStar Decided Its Workout Algorithm Needed To Shape Up

ReadWriteBody is an ongoing series in which ReadWrite covers networked fitness and the quantified self.

If Dave Grijalva were a drug dealer, I’d say he’s getting high on his own supply.

In fact, Grijalva is the CTO and cofounder of FitStar, a San Francisco-based fitness-app developer. Take one look at this tall, buff, bearded dude, and it’s pretty clear he’s partaking in the merchandise—the do-anywhere bodyweight workouts generated by FitStar Personal Trainer

In the past year, I’ve watched Grijalva go from your stereotypical doughy programmer—he used to develop video games—to a rangy code jock who switches it up from weights to yoga to rock climbing. Even then, he’s got enough energy to jam out more software when he gets home from a long day at the office.

He’s been racking up those hours putting FitStar’s code through its own makeover. Over the past six months, FitStar’s seven-person engineering team has been upgrading the algorithm that selects exercises and strings them together. On Thursday, it’s rolling out the new code to all of its users.

Grijalva sat down recently with ReadWrite to share the process FitStar went through to upgrade its app.

Leveling Up A Fitness App


Scott White, FitStar’s senior platform engineer (left), discusses a code change with CTO Dave Grijalva and VP of Operations Kristine Coco at FitStar’s headquarters.

If you’ve watched a P90X video, FitStar’s personal-training app will look broadly familiar. But instead of a standard course that’s the same for everyone, FitStar adapts each workout for you. It uses some deft digital footwork to stitch exercise clips together, so for every move, you’ll see a trainer on the screen performing the exercise with you, for exactly the same time you’re doing it.

That’s a feat of engineering in and of itself. But the real magic comes when FitStar picks your workout routine. Essentially, its software ranks you, based on an initial fitness assessment and your subsequent ratings of the toughness of FitStar workouts, and then matches you with a move of the appropriate fitness level.


ReadWrite editor Owen Thomas’s FitStar badges.

You get a little hint of this in the app’s interface: As you make progress, you get a badge showing your level in chest, core, legs, back, and other body parts. Behind the scenes, you’re getting matched up against moves on something FitStar calls The Grid.

FitStar’s training app has gotten more than a million downloads so far, and users like me have generated millions of data points. For every exercise, we tell FitStar if it’s “too easy,” “just right,” or “brutal.” Add to that other information users enter, like weight, or data FitStar’s ingesting from partner apps like MyFitnessPal. It all gets fed back into The Grid.

Grijalva’s team launched FitStar in June 2013 with the first version of The Grid. But they always wanted to improve it, particularly to add more workout variety. And now that FitStar’s planning to expand to new categories—a yoga app is coming in the fall—they realized the current algorithm wouldn’t cut it.


FitStar CTO Dave Grijalva, head of curriculum John Rafferty, and engineer Scott White discuss FitStar’s algorithms.

I Was Told There Wouldn’t Be Math

FitStar’s challenge is getting workouts to that Goldilocks level—not too hard, not too easy. The Grid is meant to match “assignments,” FitStar’s term for a specific length or number of repetitions of an exercise, to a user. Human beings vary, so it’s not going to get it exactly right every time. But the better a job it does, the better FitStar users feel about their workouts.

The mathematical challenge here is getting the points on a curve to match the ideal shape as tightly as possible.


The sharper curve shows how FitStar has improved its algorithm to better match moves to users.

At FitStar’s headquarters in San Francisco’s SoMa neighborhood, “you’ll see whiteboards that look like college math classrooms,” Grijalva says.

Grijalva’s colleagues pair the math with good old-fashioned testing. The team doesn’t just let algorithms dictate The Grid: They tweak the math based on a gut feel for when someone should, say, progress to Level 7 in Core after rocking some Superman Planks.

“It’s not uncommon to look into a meeting room and see someone doing some yoga pose or some pushups for some curriculum that we’re doing,” Grijalva says.


Workout Wednesdays at FitStar HQ.

FitStar employees also work out together once a week in Workout Wednesdays, either testing the app or trying out new areas of fitness the company is thinking about, like yoga.

I gave FitStar’s new and improved algorithm a whirl the other week. FitStar started me off with burpees—that combination pushup-and-jump move bodyweight exercisers love to hate. I cranked through a half-hour of planks, single-leg wall squats, and other tough moves. My heart-rate monitor told me I kept up an average of 127 beats per minute, with plenty of heart-racing spikes. I rated the moves “just right” down the board. It was a good workout.

My experience was typical of other testers, Grijalva told me. While the company wouldn’t share exact figures, it rolled out the new algorithm to 1 in 5 FitStar users a few weeks ago, and saw a “significant” lift in the ratings they gave workouts on a five-star scale, he said.

On the App Store, too, FitStar has been buffing up its image—its current version has a 4.9 rating, up from 4.7 across all versions.

FitStar’s Next Assignment

I found it fascinating to match my real-world experience with FitStar against Grijalva’s insights on how his team built the app. Since I started using it last year, I’ve seen subtle improvements in the quality of the workouts. That’s the result of progressive tweaks Grijalva’s team made to The Grid along the way, as well as the data I’ve been feeding the app just by working out with it—on its existing algorithm.

With its new algorithm, I think FitStar has even more potential. Like a well-trained athlete, it’s gotten stronger and more flexible.

After yoga, the company is planning to expand into other areas of fitness. FitStar has already filmed weightlifting workouts with barbells and dumbbells—it’s just waiting for the code.

I’d also like to see it make more use of the data I generate outside its app. For example, why should I have to tell FitStar whether an exercise was challenging, when it can just look at my pulse and oxygen levels? Beginners, too, might benefit by feeding in months of step-tracking information, to give FitStar a baseline picture of their fitness. Affordable, easy-to-use, mass-market wearables that can capture those signals off our bodies are coming, at which point it will just be a matter of hooking them up.

Strap me into The Grid. I’m ready. 


FitStar CTO Dave Grijalva may or may not be personally adjusting the difficulty level on ReadWrite editor-in-chief Owen Thomas’s next workout here.

Photos courtesy of FitStar

Facebook Comments

New

Rising

Popular