Inside Peloton’s AI Enablement Platform

Hi, everyone. Welcome back here to Techstrong TV. I've got an interesting guest to introduce you to today.

Not the usual tech vendor crowd that we get a lot of here on Techstrong TV. I want to introduce you to Taq Karim. Tak is the Senior Director of Engineering at, get this, Peloton.

Tak, welcome to Techstrong TV. It's great to have you on here. Hey, Alan.

Good afternoon. Thank you so much for having me on the show today. I really appreciate it.

My pleasure, man. Yeah. Glad to be here.

I almost feel like you should be on a bike or something, though, while you're doing this. But we'll let it go. Tak- Fun fact about that.

Go ahead. We've got the bike, but you should know we've got the tread and the connected rower, which is the thing I do all the time. So I've got- Yeah, the rower's great.

I've actually played with the rower. It's really... 86% of your body is worked out- Is used in that, yeah ...

is the rowing, yeah. Good for you. You probably get a better deal on it than I do, though.

I- We can talk. We'll talk. Don't give away any secrets.

You got it. Tak, you are the Senior Director of Engineering, but give people a kind of a sense of your journey. Yeah, sure.

Cool. So probably the first most important thing, I'm a New Yorker through and through. I grew up in Astoria, Queens.

I live in Jersey now, but I always- Yeah, it happens. Look, I had to move to Florida. You think you got it bad?

I'm from Queens, too. Oh, right on. Okay.

Yeah. I knew I liked you for a reason. Yeah, I actually went to St.

John's University. Oh, I noticed that. Yeah, very cool.

Yeah. Cool meeting you. So- Really?

Oh, very good. That's a good school. Yeah, cool.

Thank you. It was great. So you're in Jersey now.

I'm in Jersey now. But I grew up in the New York tech scene in the 2010s, if you will. And so, I got a degree in electrical engineering, but I've always been into trying to just hack things and play around and stuff.

So, after I graduated, I ended up in tech, and I've been basically going down the stack is the best way I would describe it. I started off as a front-end engineer, and over time, I got into the application layer, like Python, things like that, and eventually even further down into the infrastructure layer and so forth at Oracle and then here as well. But in general, I think at Peloton now, I joined in about 2021, and I've kind of been here for the past five and a half years or so.

Mainly building and optimizing high-performance stuff, and now platform stuff as well. Really? I love it.

Very cool. So would you say you're a platform engineer, or is that too confining? No, I think I've always been a platform engineer.

Okay. I just found my way to the right spot over time, if that makes sense. You migrated to where your passion lie.

Correct. I tell my own children, my own sons who are now done with school, "That's what you got to do. You've got to follow your passion and-" 100% ...

do what makes you happy with that stuff. A lot of people are watching this, though, Tak, and they're saying, "Peloton," right? " Well, one thing is, as we were formerly talking about, kidding around before, Mark Angie said it 15 years ago, every company's a software company.

100%. Software's eating the world, of course. But Peloton more than most, because the connection, Peloton took a stationary bike, and with the help of software, an app, made it ubiquitous.

Right? No one would know Peloton from any other stationary bike were it not for that technology connection. And so, I would put forth that Peloton is first and foremost a technology software company.

100%. We for sure we are. And, if I may, I'll posit it in a specific way.

So the way I describe it is we deal with very large, sometimes viral, live classes, okay? And when that happens, if you imagine what traffic might look like on any given day, it's like a sine wave, right? , low at presumably noon when everyone's at work or something like that.

But during these live events, like a Turkey Burn, which is our November kind of big event, the largest event of the year. Last year, we got about two Madison Square Gardens worth of people all working out at the same time, producing data that we then sort and stuff like that. What happens to our waveform then is it's basically the derivative goes to infinity, because you have traffic, and then boom, right?

Yeah. And it goes back up this way. That's exactly the same problem that Netflix and HBO and all these other companies have.

The only key difference is, of course, the volumes, which are two Madison Square Gardens worth and so forth. So we are at heart, especially from the platform perspective, an incredibly tech forward and tech-oriented org, because we have to be in order to provide the best experience for our members and so forth. Absolutely.

Tak, I want to kind of shift gears a little bit and talk AI a little bit, right? Okay. Now, one could easily see how Peloton, how the app, could use AI to make the user experience better, right?

Almost all the AppSec and I play with them all. I got Apple Health, I got Oura, I got this new thing, Hume- Oh, yeah ... a band that does stuff, and I just ordered something from a company, Panther, and I play with all of them because I'm crazy like that.

But actually from a geeky perspective, I like to see what they're all measuring, how they're using AI to formulate and predict stuff. One could see how Peloton would use that, but I think there's a deeper kind of thing. I want to go a little deeper with you today, and that is right down to the platform, right down to the very software.

Are you, meaning the Peloton IT team, are you guys using agentics, AI, to kind of reinvent your whole platform? Not just the user experience, but the underlying nuts and bolts of how this whole thing works. Yes, 100%.

What's been really interesting, especially in the past, I want to say six to seven months, if you recall, December 2025-ish is when the models got good, right? That's right. When stuff got real.

When stuff got real, right? Like you finally had enough to have practical agentic- Yeah ... autonomous workloads.

So what we've been really focused on once we saw that is how can we take advantage of these new capabilities of the models? Because we recognized it was going to change everything, we just didn't know how. And so our objective was really from the platform perspective, to enable access and experimentation to empower the org to be able to use AI to do their jobs more effectively or in ways that we can't even conceive, agentically, if you will.

And the thing we had to balance, the biggest challenge for us is how do we do that in a safe way, right? How do we ensure the guardrails, the security, the auditability, observability with respect to cost in order to enable the actual operators, right? So this our colleagues who are software engineers who are building all the features for our members, in addition to people in the other business units of the company, supply chain and content, so forth, marketing, and enable and empower them all.

And there's a few really cool things that we did, which I'd love to jump into with you. But I'll tell you this, the result has been like a Cambrian explosion, if you will, of tools, product ideas, and experiments that are run by engineering teams, but also non-engineering teams. And I'm happy to give some examples of that as well.

But it's been really cool to see and inspiring and just fast-paced. Really? Yeah.

So look, you can't just tease us like this, Tak. Let's jump in. Let's hear about it Okay.

So let's talk about what we did, specifically initially to enable this. So the first thing we did was we saw that everyone wanted AI. All the news cycle, all these things were occurring.

And so we said, "Okay, well, what is it that we as a platform team wanted to enable? " So as I mentioned, we want to give people access to the cutting edge models with the guardrails that we need on AI sanctioned or company sanctioned AI accounts. Okay?

Because that will ensure data sovereignty, which is a really important aspect for us, and I think probably all enterprises. Okay? And the other thing we saw was that people were already experimenting with AI, even in December or before December, if you will.

There were a whole variety of different AI harnesses people were using. There was a whole variety of different IDEs and tooling and workflow. But here's the thing, everyone wanted access to the same frontier models.

Okay? So our solution was to build a proxy. Okay?

Mm-hmm. Because the thinking was, if we have a proxy where it doesn't matter where you're coming from, as long as we can route to the places that we need you to route to to get the intelligence that you want, we're golden. Because then people can continue to operate with their own current workloads.

They can continue to have this, who knows, maybe tomorrow a whole new workflow is going to come out, and it's going to revolutionize how things work, but it's still going to be the same models, right? So as long as we can enable that to have access, we will be good. And so what we did is we built the Quarry, is the first product that we as a platform team launched.

It went viral internally and a bunch of people started adopting it and using it. And it is a system in two parts, and it's actually quite simple. So at Peloton, we are an AWS shop.

We do everything in AWS, which means that our IM, our auditing, our logging, all AWS products. So what we did is we said, okay, we built a service. And what the service does is if you're a Peloton employee and you log in through the company SSO, we are going to take your identity, which the SSO will validate for us, and we're going to mint a short-lived AWS token with the right scopes in Bedrock.

Now, I can go a bit more into why we chose Bedrock in a minute. But what this does, though- Well, if you're on AWS, it's an obvious choice, right? Because it's native, and it plugs into everything on the back end.

And the cool thing is that you no longer have model lock-in because they provide so many different types of models, including, and this is really interesting, open weight and open source models. Yeah, no, they don't discriminate with that. You're 100% correct.

Tak, I'm sitting here listening to you. Let me interject. So later this month, September 26th, we're putting on the ninth edition of what we call our DevOps Experience virtual event.

And this year it's about six or nine months ago, I picked this. Okay. Pick your AI agent.

Because I thought people would standardize, this shop is a Claude shop, this shop is a Codex shop, this shop is using whatever Perplexity or OpenClore or what have you. Funny things happened. Everyone's using these so-called harnesses.

Bedrock, at its base level, is kind of a harness, right? You got one harness, and you could plug in all your models behind it and pick which one for a specific job. And I think that's actually the winner.

The winner is I want them all , and I'll use the right one for the right task, and I want to be able to have that flexibility. I'm not going to pick one agent. And the most important thing around that is also it depends on the type of work that you're doing, too.

Yeah. Right? Planning agents versus execution agents are two very different things.

And you don't need world-class intelligence to generate code anymore. You just don't. But you definitely want something like that if you're going to plan your next big systems architecture.

And so that flexibility is incredibly powerful. And then to be able to provide at a granular level, at the user level from the company SSO perspective, that kind of access, like you have these models available, and you have these other models available. It's a very powerful concept that we're seeing already and provides us with a lot of options on how we can maximize AI enablement across the org.

I love it. You know what? This is exactly what we're seeing.

It's exactly what we're going to be talking about on the 26th. I'm going to talk to you about it when we're done with this interview. Absolutely.

Talk to me how this has been kind of accepted within-- Because at the end of the day, we're all salespeople, right? We're always selling. Always be closing, ABC.

Always be closing. And you've got to get people's buy-in, otherwise it's all for nothing. How's that going?

So here's the strategy that we took. So I have a thesis. I didn't mention before, but I actually spent a long time, about 12 years of my life, teaching on the side, coding and things like that.

I was a professor at Baruch, et cetera. And so, my thesis is that no matter where you look, there are people who are wired in the way we engineers are. The key difference is that they didn't think in their former years at all, like, "Oh, I want to go into math or engineering," and they went through some other thing.

But they exist, and they're in every single org that you can now think of. I think we can call them now AI-pilled people. They're actually revealing themselves because these are the people who care so deeply about trying to automate a system or solve a problem or go faster, that any tool that will enable them to do this, they are going to adopt eagerly.

And the beauty of it is, when you focus on enablement, like the AI enablement that we did, where anyone can just grab it without any friction or toil from our team, they will just reveal themselves. So they flock to us like, I don't know what the analogy is, but you know what I mean. Like moths to a flame, my friend.

Moths to a flame. There you go. Moths to a flame.

I was thinking something about fireflies. " I'm like, "No, no, we all have the AI. " And what we did with them is we said, "Okay, here's how you can access that stuff.

Go. " And they'll come to us through office hours and things like that and have questions or want to learn more or advance themselves. Now, the beauty of it is it's twofold, right?

One, that gives us a free roadmap of, okay, well, we have this product. How can we actually go and improve it? And these are the people who helped it go viral internally, as I mentioned before.

But the subsequent piece of it is that then they will disseminate this information, right? All we have to do is just support ways for people to do that. So we have this thing called Prompts and Coffee, where it started off as engineering, and it's just grown and grown and grown, where I think 12 months ago, people were providing presentations on like, "Oh, here's what I've learned with AI" and all this stuff.

Now you see people from all across the org talking about very specific problems that they're solving about automation tasks or data pulls. And the beauty of it is now you might have someone from marketing and you have someone else from supply chain. The problem they're solving is the same shape, just two different orgs, but now you can cross-communicate.

And more importantly, an engineer trying to teach a supply chain person how to solve a certain problem will be well-received, but a supply chain person trying to teach other supply chain persons how their problem gets solved with AI just hits way better. And we're seeing that as the proliferation of it all. The other piece of that is important, though, is the company needs to be behind it too.

And what's been great about Peloton is that our leadership, like Francis Shanahan, our CTO, probably the first AI-pilled engineer at Peloton proper, he's been providing space and support to do that too. Air cover. He's been behind.

It's been great. Yeah. So let me sum it up.

Number one, what you're finding is the best way to spread AI is peer to peer. It gets viral at that level. Yeah.

Right? Rather than top-down, bottom-up, that kind of thing. Number two, though, it is important to have that air cover from up top so that people are not inhibited to experiment, to try things, because not everything works.

We went through a similar thing here. We're not as big as Peloton, obviously, but we went through a similar thing here, and I, as the CEO, encouraged, "I don't care what you spend on credits tokens this month. " Right?

Because if they don't work, good, then we'll know it don't work. But if they work, what the heck, we'll keep going with it. And we also saw this Cambrian explosion of everyone doing this one was on OpenClore.

This one was using Perplexity Computer. Other people were on Claude. Some were in the video production team, some were in editorial, some were in market.

It was pretty amazing. Pretty amazing. And it sounds like you had a very similar kind of experience.

It's funny because this is the AI era, but to me, I think the biggest takeaway has been just how amazing humans are, right? Yeah. And how enterprising they are, and how they can help each other and be resourceful.

I will say, I think the one other thing, though, is to me, and I strongly feel this, the purpose of a platform team is really to just get out of the way. And I think we really pulled that off with Quarry and then Bureau, which we talk about in a bit if you want. And I think that just to have it available at your fingertips is such a huge plus one and acceleration factor.

And it's been amazing to see, and we're still in it, which is really cool. I love it. Tak, I'd love to talk to you more, but we are over time.

I'm actually getting a thing in my ear about another call. Got you. We will continue the conversation, but for now, thank you so much for coming on here and letting us know, and for keeping us in the know.

I appreciate it. Keep up the great work. We're going to take a break here on Techstrong TV.

We'll be right back.