top of page

Amit Aggarwal on The Robot Brains Season 2 Episode 20

 

Pieter Abbeel 

Today, we're going to talk about a topic we've never discussed on this podcast before,  designer clothes, style and artificial intelligence. Imagine swiping right on clothing items from your favorite brands that you've never seen before and having them automatically saved in your cart. Your cart spans multiple stores, but you only have to check out once. Soon, the app is suggesting brands and clothes you've never tried before, but you love them all. You're dressing better than you ever have before. The Yes, co-founded by Amit Aggarwal and Julie Bornstein, formerly of Stitch Fix, does just this and says their AI can become your best personal shopping assistant. My producer, Alice, tried it and assured me it worked really, really well. I personally tend to wear jeans and t-shirts myself, so I'll take her word for it. Amit, welcome to the show. We're so excited to have you here with us.

 

Amit Aggarwal 

Pieter, glad to be here and looking forward to the conversation today. 

 

Pieter Abbeel 

Same here. I mean, this is a very different topic from anything we've covered so far. Clothing, style, fashion. I'm really excited that we're getting to cover this with you. So now before we dove into The Yes specifically, I'd like to kind of go back a bit earlier in your life. I mean, you're a seasoned technologist. You've worked at many of the leading tech companies. How did you go from working on Google, Inktomi, Bing, Groupon to ending up working at THE YES, in fact, founding THE YES? 

 

Amit Aggarwal 

Yeah. In fact, if I think about my experience, there's a common thread of what I've done over my career, which is the way I think about it is that my career has been about using data and machine learning to build great consumer experiences. So when I was at Microsoft, I spent a lot of time on the Bing search engine. This was almost 15+ years back, but we were at that time using neural networks to do web search ranking, way before, you know, they became really mainstream. And so in that case, we're using AI to do web search ranking. At Google, we're using a lot of data to build advertiser tools. At Groupon, I was building personalization technology. So the common theme across all of these things was the idea of really leveraging data, machine learning to build personalized and relevant user experiences. And the genesis of the years, to some extent, is the idea that despite all of that work, despite all of the advances in AI and machine learning and personalization, if you think about online e-commerce today, it is pretty much one size fits all. And where we all talked about personalization, and AI for a while, it doesn't show up in the user experience as much as we would like it to. And so we started THE YES with the idea that we can solve that problem. We can reduce the overwhelm that exists in online e-commerce shopping, where you're inundated with choices and pages and pages of products that you have to go through to find out what you're looking for, by kind of bringing the idea of personalization through AI and machine learning to users. And that's why we built THE YES. 

 

Pieter Abbeel 

Now that's so interesting. I mean, what typical site would go to the shop if I go, let's say, buy some new tennis gear for myself? I'm essentially checking off a bunch of boxes like, you know, men's wear, maybe shorts. And I specifically have to narrow it down to see the things I want to see. But THE YES is very, very different, how it's set up. Can you say a little bit about when for the first time somebody visits THE YES, what is the experience like? And how does that evolve as they visit more often? 

 

Amit Aggarwal 

Yeah. So we started, we launched the first product as an iOS app. And when you download THE YES app and start using it, the first thing that you do is you go through an account creation and then you answer some questions. And these questions really give us a sense of who you are and your preferences around style and what you buy. And it's a set of questions that is strategically built to get the maximum understanding of you as a user and your preferences to train our machine learning algorithms. And once you've finished that quiz that takes, you know, maybe 10 minutes, you land on a home feed. And this is in some ways you could think of this as your store. It's as if you were entering your store. Our vision is that if we have, you know, let's say, a million, five million users, we should have five million different stores. There's no reason that when you and I go to our online store experience, we will go into the same store experience if the store is a kind of online store. There is no capital cost of building that store other than software. And so all of us should have our own store. So anyways, you complete this quiz and you land in the home feed in the home feed is your store. It is as if someone created a store just with one customer in mind, and that's you. And so that store has the sizes that would fit you. It doesn't have any, any, any products that wouldn't fit you. It is the size that you would like. It's the brands that you would want. So that's that side of the feed. And really, it's kind of reducing the overwhelm that you see in more traditional e-commerce where you are seeing everything and then you have to remove the things that you don't like. The idea of this store is that it is your store. Further and really key concept in what we're doing is this idea that the store is, won't be perfect, but it needs to be constantly learning and evolving and getting smarter and having sort of almost that conversation with you to try to understand you. And so we have this concept of yes and no on the product, which is, you know, anytime you see any piece of content on our app, you can either say, yes, I like it or no, I don't like it. And the system is getting smarter. It learns from those choices that you're making about what you like and don't like, to learn more about you to evolve its understanding of you and get smarter and get better over time, which is really key. We realize that the goal of personalization really is not to give you the perfect product every time. I think that we would be too arrogant to think that we could do that. It is more to kind of build this around you and learn from you all the time and this is what THE YES does. And so one of the important, two important concepts when we're building a store around you, it's getting smarter almost in real time as you interact with it. 

 

Pieter Abbeel 

Now it's very interesting what you're saying there, you can't always perfectly recommend because people, I imagine most people's taste will also evolve over time. And what was great for them yesterday, maybe not next week. And maybe, at least for me, if I buy jeans today, I probably want something else like a T-shirt tomorrow. So I'm curious as somebody signs in, what is some of the important information that you have to collect, such that you can correctly personalize that store? 

 

Amit Aggarwal 

Yeah, there's this where, by the way, one of our core beliefs is that in order to do great personalization, you have to go deep into the domain. I'll come back to your question. But we really believe in the idea that in order to work great personalization, you have to go deep into the domain. There's obviously techniques around more horizontal general personalization, collaborative filtering is an example where you don't really need to understand the domain deeply. You really use user behavior data to personalize. But we believe that has limits. So what we've done is gone really deep into the fashion domain to understand what matters. And there are four things that matter for fashion when people think about the buying decision. Number one is style. You talked about tennis clothes and shoes, and so you might have a style there. You buy some types and not other types, and you'll be open to, and there's going to be a set of styles that you would never buy. There's going to be a set of styles you always buy, and then there's going to be another set you’re going to be open to. So style is really important in the buying decision. Size and fit is really, really important in the buying decision. And also whether the customer keeps the product or not. Price is always, across e-commerce, an important buying decision. And then brand also plays into the buying decision. So these are the four key areas that influence the buying decision where we want to personalize. So the questions that we ask are geared towards getting an understanding of these things. And typically what we realized is we've done more and more user research, is that the way to learn about these questions is not necessarily to ask the question. For example, if you go to a quiz, you won't see a question that says, what price would you buy it? And the reason we don't ask that question is users themselves don't know what price point they want to buy at, and it changes all the time. And even if you ask that question, we wouldn't get an accurate answer and it would become out of date frequently. So we actually don't ask that question. We use other ways of learning about that question. So that's one principle that we use is, how can we get data that is easy for the user to answer and that actually stays relevant over time? So I'll give you one example, maybe to illustrate the work that we've done on how we ask the question. So, for example, think about style. We could simply ask you as a user, what would you like? What styles do you like? Do you like contemporary? Do you like modern? Do you like boho? And the challenge with that is that a lot of users don't necessarily know what they mean. They might actually, even if they know they might have a slightly different connotation associated with these words. And so that's not necessarily, that doesn't work that well. So we experimented with this idea of what if we showed them the image of the product and said, do you like this image? And you liked this product from a style perspective? And the issue is, we did user research. We noticed that users would try to answer that question, but they would actually really have a hard time understanding what aspect of style we're asking for? Are we asking about the color? Or the pattern? The silhouette of the product? What are we interested in? And so if you go to our quiz today, we actually showed three images as a cluster and said, you like this question? And these three images are actually automatically done by our machine learning algorithms. So the machine learning algorithms figure out style clusters and pick these three images to ask the user. And what that does is that when you show three images, the user's human mind is really good at focusing on what's common between them. And they say, well, okay, I understand you're asking me about color. These are white. So you asking me whether I like white or these are all yellow stripes? And so you're asking me about stripes? And so this is an example where, there's actually really, really interesting, fascinating example of how it's an example of AI, which is behind the scenes along with humans, AI that is finding clusters of styles along with a user interface innovation where we're saying we're not going to ask about a textual question or a single image. When asked about three images, because that's the most efficient way of asking it, together, working in concert to help the process of personalization. 

 

Pieter Abbeel 

That's so interesting, Amit. On the podcast, we've had many guests talk about the importance of data and having the right data, have it labeled correctly. And a great example has come back many times as self-driving cars. You need to label data that's experienced and you need to actively go collect the right data that's interesting. In your case here, it's yet a different way of achieving the right data because you're saying that one data point at a time is hard to label correctly. But once you show three images in this case, the human labeling process is going to be much more precise and you actually get the signal that you want, and the neural network can pick up on this signal and internalize what the person actually cares about. 

 

Amit Aggarwal 

Yeah, exactly. And I think you're spot on. Its labeling, how you label data is a super important thing. In this case, we're getting our users, end users to label what's important to them and then learning from that. 

 

Pieter Abbeel 

Now you talk about style, which I think is probably, it's definitely what I first think about when you say like a shopping assistant, that helps me. I mean, this should help me identify things that have the right style. But you also mentioned other things. You mentioned size. And especially on the women's side, size is often not just one number. It can be a pretty multi-dimensional concept to get the right fit that matches the person. And so I'm curious, how do you go about that? That's very complex, I imagine, to figure that out. 

 

Amit Aggarwal 

Yeah, it's a very, very complicated problem. This is an area where there's enormous opportunity. If you think about the impact of being able to clearly figure out fit is huge from a customer perspective. You know, like a lot of reasons, customers are frustrated is because something that can’t fit them and they have to return it. But in thinking even bigger, there's a huge environmental effect of this. You know, there's a lot of like a lot more broadly in the industry. Return rates can range from 30 to 40 percent, which is basically if you think about it, that's 30-40 percent of the products going to the customer and going back, without any transaction happening. That's a waste. If you could prevent that, that's going to be great for the environment as well. So it's really important, I would say, you know, maybe five, 10 percent done on that problem. The first step of solving that problem is actually really simple, which is that sizes are not consistent across brands. And so a small and one brand could be the medium on another brand. And so, you know, this inconsistency, that by itself is actually an important problem to solve because a user might understand this style, the size and one brand. But when they start buying a new brand, which is, by the way, one of the things that we try to do is introduce them to new brands. They don't understand the size there. And so that's a problem that we've solved with data where we kind of learn from our data and map different brands to a common taxonomy of sizes that we've built. And that really helps. You know, when I was talking initially, I said, well, yes, the store only contains things that would fit you, so you automatically do that behind the scenes. We understand what size fits you, let's say in one brand, we can automatically map it to all the other brands and only show you the things that would fit you in any brand across our entire platform. So, that's an important problem solved. Beyond that, I actually think there's a lot of opportunity in going deeper, you know, size is one thing, which is what size would fit you the best. But going further, the problem really is whether it would fit you well. It might be the best size for you, but doesn't mean it'll fit you well. And that is a much harder problem. There are data challenges with that problem, which is how do you actually get the exact measurement of products? There's no clean data around it. You know, if you order every product and measure it, that's too expensive. That's not scalable. That's super expensive. So there's this unsolved data challenge there. How do you get the size of the user accurately? There's a lot of unsolved AI and user interface challenges. Users typically don't want to buy a $500 camera and then take a picture of themselves to get their size. That's not a user interface that works. So is there a simpler user interface and a simpler AI technology behind it that can really seamlessly without much friction from the user, get an understanding of their size? I think that's a problem that we think about. So I would say, you know, there's a lot, like it's a multifaceted problem and potentially a big area of impact for AI. 

 

Pieter Abbeel 

Now, as I'm thinking about everything you described so far, I'm trying to wrap my head around what happens as somebody visits the store. So they've come in, they've given feedback on a few prompts about things they like, don't like. And then somehow there's a neural network behind the scenes that now makes a decision. And can you say a bit more? What is that neural network doing at that point? What is it taking in and what is it then producing? 

 

Amit Aggarwal 

Yeah, the neural network is taking in all of your answers. So for example, if you had gone into the app and said yes to 10 products and no to 10 products, it can then use that data to predict what else you would like. Underlying that really high level description of the problem, there's a number of technical problems that are being solved. Number one technical problem that we spend a lot of time on is that we try to create a style representation of every product, kind of a style DNA of every product. Think of it in machine learning terms, embedding of every product that is built to represent, that embedding is of style to represent the style of the product. And it is kind of a lot of work there that we do. We're using computer vision because a lot of styles are visual. But once you have that embedding, then you can teach a neural net what embeddings this user would like. The key challenge there, of course, the biggest challenge is that the data is small. We're not talking about thousands of examples from the user and the data is small, and it's actually a little biased because, you know, users are not, it's not like they're getting a random sample and they're saying yes and no to that. So they're saying yes and no to what they've seen. So that's a big area where we've spent a lot of time thinking about how to improve that. Obviously, the embedding helps there, because of the constraints and the learning problem. Obviously, we can use data across users to increase the amount of the data. But really, what we're trying to do is, with a small number of data points from the user, learn their style. 

 

Pieter Abbeel 

So I'm envisioning every user gets turned into essentially an embedding that captures exactly their style, their size, everything. And now these embeddings, am I hearing this correctly? They get matched up with embeddings of clothing. You can take in an image of clothing and maybe some other specifications, and a neural network can turn that into an embedding of its own. And then you can try to find matches. 

 

Amit Aggarwal 

Yeah, exactly right. I think that's a really good, high level description of what we do. 

 

Pieter Abbeel 

Now when you take this in. Of course, there's many ways to take things in, for the user you take in, the shopper essentially swipes to provide preferences. Does this shopper ever also maybe take a picture of themselves? Is there anything else they can do to help this process? 

 

Amit Aggarwal 

Yeah. Not today, not today. We don't, there's no such functionality in the app today that does that. But you know, that's maybe a good idea for, we obviously have to be kind of mindful of what users want to do, would they want to do that, whether they have concerns around that. And one of the things that we've talked about is typically users have pictures on their phone and you can now bring machine learning models on the phone. And so there's potentially a way where those models could be run on the phone without actually having to upload the pictures anywhere. And with the, in some ways find that embedding for the user on the phone itself. 

 

Pieter Abbeel 

It could even find pictures of the user wearing things they bought in the past that they like and analyze that. Now on the other side, you have to also understand in principle everything that's available, right? I mean, the best experience, the best shopping assistant should look at all the clothing that's being produced anywhere in the world and find the best things. How do you go about that? 

 

Amit Aggarwal 

Yeah. So core to our business model is that idea that we need a big selection. And so our business model is based on what's called dropship, which basically means that we don't, we don't carry any inventory, we don't have any warehouses. We never touch products, but we integrate with our brand partners using technology and tap into their inventory. And what would that allow us to do is that it allows us to have a bigger catalog over time because there isn't capital expenditure for us to onboard new products or new brands. And so that allows us to give our customers access to a much larger catalog. So, even today we have a pretty big, big catalog, but the question you ask is a great one. And it's really fundamental to our business model. And so over time, because we're integrating user technology, we want to have a pretty broad and maybe the biggest brand selection out there on our platform. 

 

Pieter Abbeel 

Now you work with all these brands and the dropship process means they will do the shipping for you. And I imagine returns processing. But how do you then take their clothing into your system? 

 

Amit Aggarwal 

So we spent a lot of time building a technology platform to do that, and we use a wide variety of different ways. In some cases, our brand platforms will give us a feed of data. The cases we can tap into APIs that their e-commerce platform has. So if a brand is on something like a Shopify, then we can tap into that through APIs. And we've also built some pretty deep crawl technology that can go crawl the website of the brand and to ingest all of the products they have along with the data. Pretty much the only thing we need is the image of the product and the description of the product and the price, of course. So those are the three main things that we need for any product. We don't expect any taxonomy. I mean, we don't expect any attributes because pretty much everything else we can extract using AI and machine learning. That's actually, by the way, a big application of AI and machine learning is that we’re just given a product image and product description. We can pretty much infer about a thousand attributes about that product using AI. So we build that platform that crawls, gets data from different sources, feeds, APIs, crawl and then a layer on top of that that normalizes everything automatically. 

 

Pieter Abbeel 

Mm hmm. That's so interesting because I mean, I was thinking about internet search engines crawling the web to find all the latest. But you're essentially you have an engine that crawls all stores, all clothing stores, online clothing stores, 

 

Amit Aggarwal 

Not all of them, the ones that we partner with. So we're not a search engine in that sense. Everything that we crawl or sell on our platform, we have explicit partnership with that brand. 

 

Pieter Abbeel 

Well, but yeah, I imagine that's a good idea for pretty much any brand to partner with you if people shop through you. Otherwise, there is no exposure to their brand, right? 

 

Amit Aggarwal 

Yes, that's the goal long-term.

 

Pieter Abbeel 

Now it's been a year since you've been live, actually a little over a year at this point, right? Any lessons you've learned from how you thought it's going to go and then what actually happened when you went live? 

 

Amit Aggarwal 

Well, I think some things that went as we planned or hoped was number one, I would say this validation of the basic premise of the product and the basic pain point. Which is the basic pain point that we started off with is customers are overwhelmed with choice. They don't have a place that understands them and gets smarter about what they want or what type. And on that premise, we have customers who say they will have a hard time going back to the old way of shopping. They're so enthusiastic. They come back, they look at stuff very often. The repeat rate is high, really high for those customers. And the feedback that we get is I can't imagine how I would shop without this. So I think that basic pain point exists and that has been validated. The other thing that we hoped and kind of turned out to be true is that customers would be willing to share information with us and kind of give their feedback and engage. And so, for example, I talked about this, yes no functionality we use is going to give us feedback. So when we started building the product, it wasn't obvious that users would engage with it. It turns out that, from the very start, users have been engaging. They love it. They love the idea of, it almost feels like it's an entertaining thing for them to get that feedback. Especially into the like, good if they see the impact of it, if they feel like there is ROI. So that has been really good. We've got a lot of data from our customers. When we started we hoped that would happen. We thought that would happen. But there was always a risk, there's not much personalization we can do if you don't get the data. The third thing that I would say is that, one thing that we've learned is that personalization is super important. But the core fundamentals of e-commerce, which is having a broad selection and having good pricing are really important as well. Users, rightly so, care about going to a place where they can find things and they can be sure that they have the best pricing. So those are areas where we've sort of we're thinking more about and doing more and in fact bringing AI and machine learning even to those areas. As I mentioned we make it really seamless to integrate brands so that we can carry more brands. And that's an area where we're using AI a lot. Even at the price we offer, we're not a discount retailer, but we want to give them the best price possible. So those are areas that we're investing more in. 

 

Pieter Abbeel 

Who are your kind of most prevalent or power users? 

 

Amit Aggarwal 

Yeah, if you think about this domain in the buyers, there's my co-founder Julie Bornstein who you mentioned her in the beginning, has been in this domain for a long time. She talks about four kind of groups of customers. There is a not interested customer. And there's the easygoing customer, who you know is really not so interested, is not aware of brands, is very easy going about what they buy. And then there's a fashion follower. So the fashion follower customer is someone who knows about brands. You know what they are, but at the same time are looking for inspiration from really the fourth group, which is the fashionista. So the fashionista is the one who really knows what she's looking for and is really aware of all the brands, the styles, who's aware of what her style is, and is aware of what exactly she's looking for. Those are the four groups if you were to roughly segment the whole market of users. And we really go after the fashion follower and the fashionista. So our platform is not so much for the not interested or the easy going, but more for the fashion follower, the user who's really aware of the brands she wants to wear. But it's looking for more inspiration, more ideas, new brands that are coming up that she's not aware of. And then the fashionista who really knows what she's looking for, but needs help with search and searching for exactly what that is. 

 

Pieter Abbeel 

It seems from the other side for brands, you could play a very big role for new brands to be discovered, right? 

 

Amit Aggarwal 

I think you're absolutely right. I think in two ways. If you take that analogy of, we're building a store for every user. And so let's say we have 10 million users. We have 10 million stores. Every store is different. A new brand comes in. I guess, if we can help them be placed in the right stores, in the stores of the users who would like to see that kind of a new brand, that's value both for the user because they get to discover and for the brand, because they get to discover their users. That's really kind of the true promise of our marketplace in that way. 

 

Pieter Abbeel 

Very interesting. You use that word, discover, because I mean, I'm obviously not the biggest shopper myself, but when I talk with people who go buy new clothes and so forth and they're excited, often part of it is that they discovered something new. And are you worried at all that you make it too easy so people cannot actually, you know, have the discovery experience? 

 

Amit Aggarwal 

Actually discovery is a big part of our algorithm. So the goal of our personalization algorithm or the technology that we've built is not to show the product that the user would buy, right away. It is to help them in their journey of looking better, and that journey does include discovery. And so your traditional personalization algorithms do have this issue that they really narrowed down into the things that you've done before. Especially the ones that are based on user behavior data. They really kind of focus on the things that you've done before. And so users get bogged down in that place.

 

Pieter Abbeel 

You get put into your own bubble.

 

Amit Aggarwal 

Into your own bubble. There are two things that we do and do to prevent that. One is that a big part of the algorithm is discovery. So we're not necessarily trying to say what product would you buy? To some extent, we're trying to say what product you do not want to see and we can remove that. And then among the risks, we can provide you diversity and inspiration. That's one. Number two, I think this idea of yes-ing and no-ing a product is important because what it does is that we can explore more. We can be more aggressive about showing things that you might not like, but that's okay because you have this way of telling us. In most experiences you think about the reason you can't explore much is because there's no real time explicit clear feedback from the user. We have real time explicit feedback from the user so we can explore. And then if the user doesn't like it, we will adjust right away and fix it? 

 

Pieter Abbeel 

Now, when I think about this, I mean, as I think you know, I work a lot in reinforcement learning, right? And part of what you're describing here also reminds me of reinforcement learning, the exploration process and then finding the right thing and essentially guiding the shopper through a journey rather than a one time decision. And I know reinforcement is still a much earlier stage as machine learning technology wise, then supervised learning. But I'm curious, is there already some reinforcement learning happening under the hood here? 

 

Amit Aggarwal 

Honestly, not a lot. We're in the early stages of exploring it. I think it's a good fit for what we do because you're absolutely right. We think of this as a journey, and reinforcement learning is a great way for us, a great technique for us to explore using. I think we're getting at the point where we have the data to do it. So I would say it's a great direction for us and we're in the early stages of exploring that. 

 

Pieter Abbeel 

Well, I'll be excited to hear more about that when that starts happening in full force. Now I'm curious, THE YES is what you've been working on the last few years, but before that you were at Groupon, which seems to have some connections. You were at BloomReach, which I think must have some connection. So I'm curious, are there some lessons you learned there that you brought with you as you're building THE YES? 

 

Amit Aggarwal 

From both BloomReach and Groupon, one key lesson was personalization is important. Understanding the users is important and giving them a relevant experience that they feel is relevant to them is important. I mean, we saw that with Groupon that the personalization technology that we built had a huge impact. And then BloomReach was more of a B2B company. So we built technology for other brands and retailers, and I work with a lot of them. So this idea of personalization and providing an experience that's tailored to the user and a dynamic experience. Both of those experiences reinforce that. I would say one thing that I learned, though on the other side, is that it's really hard to fundamentally change the user experience if you are an established business. It's kind of a very risky thing for an established brand or any business to fundamentally rethink how to do things, to fundamentally change the user experience. And that limits how much you can do that. Number two, I learned that this idea of thinking about AI as a black box where you have an AI algorithm that will come up with magic is, I think it has limitations because it sometimes oversells the potential of the algorithm. But more importantly, I think users are not looking for that perfect recommendation. They're actually looking to have a conversation. They want to understand the decision they were making. They wanted to have an influence on that decision, and they want to feel like what they have to offer is being used. And so this idea that you want to build an AI algorithm that does great behind the scenes is probably not the right direction. And what we've tried to do is sort of combine the UI and the AI and have the work very closely. And I gave maybe a couple of examples of that. The third one is, I think, domain knowledge is important. In order to solve some of these problems, you really have to bring domain expertise into train the machine or think about the problem. And that's important to solve some of those problems. And that's something as well that we've brought to THE YES. That's the reason we're starting with the domain that we are in. We might go into other domains where, as we go into other domains, it's going to be with a very deep domain focus. So those are some of the learnings and how we've applied. 

 

Pieter Abbeel 

Now, as you think about, you're very focused now. But now what if you zoom out a little bit and think about the future of fashion, clothing, retail and the role of AI and technology. How do you envision this will be five years from now? 10 years from now? 

 

Amit Aggarwal 

I think of AI as software. Like it'll be what software is today, which is pretty much every business, every part of that business is impacted by software and improved by software. I think I think of AI as the same, which is, we're really, for example, right now focused on applications of the AI to search and personalization. And that's been a core focus. And, I think there's enormous amounts of work that we will continue doing in that space with personalization, size and fit, all of all of those areas eventually AR, VR. But, pretty much we want AI to impact every part of our business with its customer service, whether it's how we integrate with brands. I talked about, when we integrate with brands we use AI to normalize all of that data. We're going to continue doing that so that there's less and less manual involvement needed there. Same thing with customer service, I think there's a huge opportunity to make customer service much better by leveraging AI. There's a huge opportunity, for example, I mentioned returns as a problem in the industry. There's a huge opportunity to apply AI to bring that return rate down substantially. So I view AI and the applications of AI very broadly, and I feel like in five years we'll not think of it as something we apply to a specific thing. It's actually going to be in some way shape or form be used in every aspect of our business. 

 

Pieter Abbeel 

Now that's so interesting. And it's a very broad scope. The one thing I'm also curious about, especially with my Covariant hat on being very active in warehousing and so forth, is that side of things. And you touched upon returns, like reducing returns, the amount of returns processing. I'm wondering, even beyond if we think about manufacturing and logistics all integrated. And do you see any trend where things could be tailor made effectively, where it can be tailor made for a specific person, just in time? And then that's made exactly for them. I mean, obviously today, tailor made clothing is very expensive, but do you see a path to that also? 

 

Amit Aggarwal 

It's a fascinating idea, really fascinating. We haven't thought that much about the, you know, logistics side of the things in the warehouse and manufacturing. We rely on our brand partners for the most part. But yeah, I really think that's a fascinating idea. 

 

Pieter Abbeel 

We've gone through a very special time or kind of still in this special time with the COVID pandemic. And it obviously affected many people in many, many ways, including a lot of people who started doing a lot more of their shopping online. And of course, THE YES is an online platform. And so I'm curious, how did you experience COVID-19 from a business perspective? What did you see happen? 

 

Amit Aggarwal 

Yeah. We were supposed to launch in March of 2020, right around the time we locked down. And so obviously, our first decision was to delay that launch because it was an uncertain time and probably not the best time, given the uncertainty in what was happening outside to launch a new shopping platform. And I think we eventually launched in 2020. It just gave us much more time to really fine tune our user experience in our product. So that was a good opportunity for us to take a step back and use the time to fine tune it. There are two things that happened with COVID. One is I actually think it really accelerated the move towards online e-commerce. And there's plenty of data that shows that, maybe accelerated by three to five years. It was already a pretty clear trend towards online e-commerce. But  with COVID it has accelerated. And so, that's good for us. We have a bet on online, THE YES is an online shopping commerce. And that's, we've always believed that's the trend, even for categories like clothing where the penetration of online e-commerce is lower than other categories. And we always believe that it's going to go in that direction. If anything, COVID has accelerated that. In the short term, during COVID, there was certain categories that did better during COVID, like loungewear and so on. But we're seeing sort of, since the vaccine, we're seeing things getting back to normal and we're seeing a growth in our business and kind of, the normal diversity of categories that users are going back to. So really, at a high level, it's just more of an exploration of that trend that we see now. 

 

Pieter Abbeel 

You raised $30 million in 2019. And any plans on that front and any directions you're growing the team into? 

 

Amit Aggarwal 

Yeah, so you're right. We raised 30 million. And we launched our iOS app last year in May. As I said, we spend a lot of time fine tuning the product. We reached a point where we're really happy with our repeat rate into the engagement of the app. Earlier this year, we launched the website. So we’re trying to bring that same technology and same experience with the website to cover a broader set of users. Not everyone is on an iPhone. And so that has gone really well. I would say our business has grown really, really strongly in the last year, since the beginning of this year, and the website has done really well. The app that we launched continues to do well. Users are engaged. Our engagement metrics are good, and revenue is growing really solidly. So we're really focused on that. And at some point we'll raise it up a round. 

 

Pieter Abbeel 

Now you're a founder of THE YES and many of our audience, they're always curious about, how did you decide to start your own company? And what are some of the key lessons and takeaways you'd like to share for other people thinking about starting their own companies? 

 

Amit Aggarwal 

I've worked at a number of companies. I had worked at big companies, the Googles of the world. And I worked in medium sized companies like BloomReach, which I would say was a startup, pretty small when I started. And through that journey, I realized that I love building. I love building more than scaling. And what's better than to start a company to build? If you're a builder and you like building, there's nothing more fascinating and more fun than building something from scratch where you have nothing but an idea and you're building a product, you're building that technology with zero lines of code. This idea that nothing existed. And now you have this new company, platform, product is really fascinating and it's a fun journey and nothing like that with respect to learning. And so that's really the reason I wanted to start a company was to build something from scratch. The only other thing I had in mind was that the area I work in should be a huge market. Your market is important. It should be about technology and product innovation. There's many ways you can start a company, including process innovation or business innovation. But, being a technologist, I cared about being at a place that was about technology. And the team, that we could build a great team because at the end of the day, this sort of journey is uncertain and there's no certainty about the direction where you need a very good team. And so do you have the ability? Are you uniquely positioned to create that team for this domain in this problem? Those are the three important things that I get. And I think, kind of the lessons that I've learned is that it's a lot of fun. There's nothing like it. No amount of building a new business within a big company, people say that's a startup within a big company. There's nothing like a startup within a big company. A startup is very different and it's been a lot of fun for me. Things that always don't shape the way you want them to. They always take longer than you think they will, and that's okay. You just have to have the resilience and the right team and the right conviction about the idea to keep going. 

 

Pieter Abbeel 

Well, those are some great lessons, and also, I mean, clearly, it's going really well, so you made the right decision here. It's very exciting to see how THE YES is coming together. 

 

Amit Aggarwal 

One thing I would add is I've been in technology for a while, and the pace of innovation that's happening right now is mind boggling. And one of my biggest mistakes previously has been not keeping up with the innovation. And I actually think this is a time of unprecedented innovation that we're seeing in all kinds of fields, but AI and machine learning and the adoption of technology across companies is just at an unprecedented pace. The willingness to adopt technology, the willingness to change the way of thinking is unprecedented. I am really excited about where we are with technology. There's never been a better time, even in my career of the pace of advancement. 

 

Pieter Abbeel 

It is a really exciting time. Yes, especially in AI, things are moving so fast the last five years and seems only faster every year. It's amazing. Well Amit, thank you so much. This has been an absolutely wonderful conversation. Thanks for joining us. 

 

Amit Aggarwal 

Thanks so much. I really enjoyed it. 

bottom of page