Julie Bornstein: Building the Future of Fashion with AI

0:28Hey, everyone, and welcome to Generative Now. she's back with her latest company, Daydream, a bold new startup that's looking to reinvent how we discover fashion using AI. We talk about the lessons she's learned in her many years in the e-commerce and fashion spaces, what it's like being a second-time founder, and her hopes for Daydream. So let's get into it. Hey, Julie. Hi, Mike. Thanks so much for doing this. Really, really appreciate it. I've been looking forward to it. Me too. So I have so much I want to talk to you about. And I'm sure we're going to go back, jump back and forth between what you're doing now and your very sort of amazing experience, storied experience.

1:06But before we do anything, tell us about Daydream. Like give us the high level pitch for Daydream. Daydream is basically a fashion search engine that leverages AI to be able to interact as you would if you were in a store. So I think we've all, everyone who's been shopping online for the last 20 years very well knows the difference between being able to go and talk to someone and find what you're looking for versus in a store versus needing to be very specific around the taxonomy of the brand site and knowing how to find something on the web. And so with the advent of LLMs, we now have the ability to have sort of real language interaction with consumers and understand what their general need is to find what they're looking for.

1:56So we are working with all the fashion brands in the world that are real brands. The consumer can come on and ask a question, tell us what they're looking for, use regular language. We got one recent request that said, I'm looking for a revenge dress for a wedding in Paris in the vibe of salt burn. And so, yeah, you can, you know, you can ask anything you want. The agent will help you find what you're looking for. Once you find it, you can save it into a collection. You can share it with friends, get their advice and feedback and click out to buy on the brand or retailer site. Amazing. You know, when I saw this company get announced, you know, when I understood or at least at a high level what I believed it did, immediately I was like, OK, this makes so much sense for all the reasons you just mentioned.

2:44But the other thing that just made so much sense to me was there's a very common through line between your previous experiences and this, you know, I think about Stitch Fix where you were COO. And at the time I remember thinking, wow, because I remember being a customer of Stitch Fix at some point. I'm like, wow, this is simplifying and making it easier and more possible for me to shop without, you know, wasting my time in stores or navigating online stores. And then, you know, same thing with the yes, AI for helping people discover and help shop. It feels like there's a theme here. Maybe, you know, say a little bit about like kind of those experiences and how they helped inspire what you're now doing with Daydream.

3:28So when I was a kid growing up in the 80s, I would get 17 Magazine, find a product and go into the mall and try and find it. So I feel like I've been working on search for a very long time. When Amazon first launched, I remember literally like watching the site for the first time. It was the summer of 1996. And I was like, oh, this is the future. This is going to make shopping for fashion so much easier. And so I actually spent about six months trying to convince Dan Nordstrom to hire me at Nordstrom.com because I was in Seattle. My husband was at Amazon, actually. And so we were sort of there for a while.

4:03And it was they had announced that they were going to launch this big web business. And so, you know, I would say everything that I've worked on is definitely been leading up to this. And whether it was at Nordstrom building the first generation of e-commerce and then going, I was at Urban Outfitters at Sephora and then Stitch Fix. all of those were sort of steps towards what I ultimately wanted to be able to create. And in each case, I was sort of chasing technology, like as it emerged to be able to do things that made shopping easier. And so the idea for Daydream was an idea I had really when I was back at Nordstrom.com in like, you know, the year 2000.

4:46And we just were so far technology-wise from being able to build this. But I dreamed of not having to shoot, you know, samples of products, of not having to show everyone the exact same results based on what they searched, of being able to use natural language search. So, you know, we've finally gotten there. And this is very exciting for someone whose obsession is online shopping. Yeah. When I was thinking about your trajectory and your career, it does seem like all of the key experiences and, you know, especially, you know, being a technology investor, focusing a bit more, I would say, on Stitch Fix and the Yes.

5:23Like, it seemed like those two moves, and of course, Daydream, as you just said, sort of were chasing a technology. Like, talk us through the technology that Stitch Fix was chasing and the Yes was chasing and how that compares to Daydream. Yeah. So I'll go back just even a little bit before then, because so at Nordstrom, the, you know, we were trying to figure out how to sell things online. So the, the, it was the, it was the early era, you know, now the problem is overwhelm and there's too much online. And so the problem we're trying to solve is helping you find the right thing. But, you know, so we were focused on sort of very basic things like how do you get basic search to work?

5:58It was all keyword search. How do you create categories that people understand? And how do you shoot, you know, things and get them in a warehouse and get them set? By the time I got to Sephora, there were, you know, I think the big technology changes were mobile became a thing, social became a thing. Interestingly, we were doing reviews with Bizarre Voice, and we really wanted to have real-time Q &A about a product, and they didn't have that capability. They were like, we don't know what you're talking about. That doesn't exist. So we ended up working with a gaming company or a company that worked with a lot of gaming companies called Lithium to create sort of our first conversational platform.

6:36So we created something called Beauty Talk. And that was kind of based on what we had seen happening on Facebook. People would ask questions and then other people. We thought we were going to have to be the moderator to answer them, but we realized we didn't. But then what happened is the question would come and go and you couldn't get back to it. And we really wanted to create a repository of all the questions you might have around the beauty space and products, whether they were sold at Sephora or not. And so, you know, I would say in the era that I was at Sephora, which was 07 to 2015, it was sort of really getting mobile as this bridge to, you know, web and store and getting sort of these social interactions working.

7:16At Stitch Fix, what I was really interested in was this idea of using data science. And we had had some people who were at Netflix doing some early things, a guy named Eric Colson, who's great and had built some of the early algorithms at Netflix. And so it was really how do you sort of understand a consumer and serve them up, you know, the right relevant product for them. But because the model wasn't real time, it was, you know, we basically had a profile on each user and then the stylus packed a box and it was shipped to the consumer. But I loved learning about how to use AI in the early days for it's really, you know, sort of a combination of machine learning and computer vision to like figure out what the right recommendations were.

8:08But the model was somewhat limited because we owned our own inventory. And so you could only serve people well who happened to map to that inventory. And then so sort of through, you know, the early AI and ML days, we decided to build the yes because we wanted to build something that was really geared towards more of a fashion customer. So Stitch Fix really served the non-fashion customer quite well. It was a treat to get a box of items that look good and fit you and tended to be people who didn't know brands and weren't super picky about what they wanted from a fashion perspective. And so I started the yes after that because I felt like I really want to build a fashion recommendation engine for people who are engaged in the category.

8:52And so we were, you know, that was the sort of genesis of that. And, you know, I would say then we were acquired by Pinterest. We'll go back to that later. And during that time, ChatGPT launched. And, you know, the sort of understanding of an LLM, a large language model, and what it could do kind of really transformed the way I started thinking about shopping and what was possible and led to starting Daydream. Yeah, that's super, super interesting. You know, some of the other things that come to mind for me, I don't know if there were things that, you know, were important to you at the time, but with Stitch Fix, you know, it almost seems like a lot of the innovations and sort of like supply chain of e-commerce and sort of efficiency of sort of global shipping and operations, that must have been a bit of a tailwind for you.

9:39And then with the yes, the thing that jumps out at me is it feels like when you started it, and I want to get into the timing of this a little bit, it almost feels like the personalization wave, right? You know, Spotify personalization, TikTok algorithm, like that sort of thing, which I think per your point is much different than say LLMs, but also very, very important and very, very relevant for the time at which these products existed. Were those important factors for these two businesses? Yeah, super important. Yeah. And I think that, I mean, that was the whole idea behind Stitch Fix was personalization.

10:18And it's funny because we had done a lot of interesting things at Sephora around understanding the user and targeting marketing based on their past purchases and their brand affinity. But what I loved about Stitch Fix is we were using that information real time to make purchase decisions or product decisions. And so to me, it was really exciting to go from sort of analyzing data and using that data for marketing purposes to real time use of the data to help make recommendations for the user. And so that really, I would say Spotify and Pandora, we had employees who had worked at both of those places at Stitch Fix.

11:03And I always loved hearing how they built and what their journey looked like from sort of their first version of the product, where they basically had like, you know, 25 professional musicians who were keywording every song so that they understood the dimensions of a song and could make recommendations based on that to sort of using different data points over time for their recommendations. Yeah, that must have been super cool. One of the things, going back to the yes really quickly, one of the things I noticed, and not that this has anything to do with AI, but it must have been really fascinating nonetheless, is this company was started just before and was acquired just after the pandemic.

11:44And I was operating during that time and through that time and obviously know so many other founders and companies that were. And that was just a really, really crazy time for company building. I'm curious just to hear that perspective of what it was like to operate and exit throughout such a challenging moment for startups. We started in May of 2018, yes. And our plan was to launch in March of 2020. So our launch date was literally set for March 20th, 2020. And so March 12th came and everyone's like, what's happening? And of course, as most entrepreneurs, I'm like the ultimate optimist. And so I'm like, oh, this is going to pass in three weeks.

12:24Let's just hang tight. And we'll, you know, see, we'll talk next week, then we'll talk the week after and we'll decide. So obviously, a few weeks in, we realized this wasn't going to go away anytime soon. And we realized that we had to launch during COVID. And it felt so weird because, you know, it just like fashion was the last thing people were worried about. I mean, we were in this like global crisis and it just felt really odd. And then I think that what happened was, you know, by kind of like late April, early May, we were all kind of like, this thing isn't going away anytime soon. And we do need a little, you know, play in our lives.

13:02And so let's lean into the fact that this is a fun way to pass time as we think about going to launch. Yeah. And so sort of like, you know, the yes was not dissimilar from Daydream, my new company, and that you can sort of save the things that you like and help to train the model to get to know you better. And so we found that a lot of people were just enjoying yesing and knowing products and getting their sort of style better understood. And so we kind of leaned into that and we ended up launching in May. Um, we got lucky because we launched about, I think a week or two before the George Floyd incident.

13:39Um, at which point everything kind of went dark again. Um, I don't know how much you remember that, but like, you know, no companies were doing any advertising for weeks after that. There was just this like darkness that sort of, and heaviness that came over everything. Um, but we sort of happened to like get out in the little window before that. And then it was just, you know, I would say we had no business we had to comp. So starting in kind of a slow time wasn't a bad thing for us. And it gave us time to sort of improve and learn on the product. And then, you know, I think as everyone readjusted to the fact that this was not going away soon and they needed to, any purchases they wanted to make needed to go online, You know, it gave became an opportunity for the online business to really sort of grow and scale across all businesses.

14:33So that was kind of the start. It was it was definitely hard to go from being all in the office to everyone being remote. But I felt lucky that we had we had about 45 people on the team. And I looked at all my friends who were in former businesses I was in with hundreds and thousands of employees. And I really felt for them having to manage and motivate huge teams remotely. I think that would have been much trickier than having a small, tight team that could stay connected. Yeah, that makes total sense. And, you know, my startup journey was similar-ish timing. We sold to Spotify right before the pandemic.

15:07And, you know, the main thing that I remember being really challenging, and I'm curious if you had a similar experience with the Yes, was the integration, right? You obviously, during this period, you sold to Pinterest. I think that was in 2022. Yeah. You know, the pandemic was, I guess you're sort of coming out of it at that point. But what's that integration like? Yeah. At that moment where everyone is like trying to figure out how to work in this sort of post pandemic world. Yeah. Yeah. Well, so the story with Pinterest is that the team had reached out to me. I had known Ben Silberman, who was the founder and CEO at the time of Pinterest, and he was sort of watching our development.

15:43And he always felt like Pinterest should be in e-commerce and had sort of struggled to find the right team to build that. He had a lot of people who came from the ads world. And so his idea and his vision was bring the team over, use the technology, and start to use that as the foundation to build this shopping experience on Pinterest. And so the acquisition process was a little bit of a nail-biter because the market was kind of in a bit of a free fall during that time. So we started talking in February of 2022, and the deal closed in June of 2022. But Ben was amazing and, you know, very clear on his vision and really stuck to, you know, what he wanted to do.

16:31And we got the deal done, and it was very exciting for the team. It was a big win for everyone. And then, as happens in this world, Ben announced that he was stepping down a few weeks or a few months, sorry, into the timeline of our integration. I tried to be as helpful as I could. We sort of figured out the right roles for the team within Pinterest. And I stayed for a while to help with that. And then as I looked up to think about what was next, ChatGPT had launched. And so all the possibilities were really exciting. How soon after that do you say to yourself, OK, I have something else I want to do?

17:11Obviously, that thing turns out to be Daydream. So I stayed on as an advisor to Pinterest until the summer of 23. And I would say that spring, I started to talk with some people about Daydream and really just get a sense for the idea. And if there was interest for some co-founders to join me and started to talk to some investors. And then we basically ended up raising that fall. So the fall of 23. And building pretty quickly a good alpha product that we started testing in the market in the beginning of 24. So obviously, LLM is the sort of big, inspiring thing. But as we talked about earlier in the conversation, so many elements from Stitch Fix and I'm assuming the yes probably went into that initial idea.

18:00Like, you know, talk to us about sort of the inspiration and how that all led to that initial moment of sort of inception for Daydream. Yeah, I would say that the thing that I had been interested in my whole career from was search and more sort of specifically natural language search. And it was really hard, even if you built the tools to be able to do it, consumers were not trained to search that way. And so at the ES, we did some testing. You know, if we tell you you can put anything into the search bar, you know, we can understand it and react to it. It just the consumer wasn't trained to do that.

18:43And search and fashion in particular has always been really bad. And so the idea that now suddenly because of ChatGPT, there's this opening of this opportunity for consumers to be retrained and to understand how to interact with an agent and ask for something in natural language became really important because it's really hard to change consumer behavior when it's, you know, so embedded in the way people interact. And so you need something as big and broad as like ChatGPT coming on the market to sort of say there's a different way that you can start to search and ask questions of the web and, you know, be able to get information back.

19:25So I think that was kind of the first aha for me is, you know, not only do we now have this model that we can leverage to build an application, but we also have this training ground that they're doing with consumers that help them understand you can kind of ask for anything you want. And we know how to translate that into a good search. Yeah, so it almost sounds like the model providers, OpenAI, Anthropic, Google, to some extent, it sounds like what you're saying is they're almost training the consumer how to use Daydream or how to use products like Daydream. And I imagine that's going to be a huge tailwind for the product.

20:04We haven't seen the product yet. It's pre-launch. But I guess maybe help us imagine it a little bit. I'm picturing natural language search, obviously. you know, before we started recording, you gave some examples of like things people, or maybe it was in the beginning of the interview, you gave some examples of some of the things people are searching for. Give us a little more sort of flavor of what we can expect from the product. I think the UI interaction is going to change over the next few years. And I'm really excited to experiment and play with it. I would say what we've built is kind of a bridge to whatever the future is.

20:38So you can start with a query. I do think that having a search box versus like a conversation box, that small change actually really helps the consumer figure out, oh, this is more than just a one-line search. I can tell you things. We also show examples of what other people are asking for. It's really fun to see what other prompts people are using, And it also just helps give you examples of what you can prompt with. You can also start by uploading a photo or by voicing, you know, your request in. And it's funny when people start to voice, they get much more conversational. And so there's a lot more detail.

21:16So, you know, you may, if you're typing it, you might say like a red dress with long sleeves, three quarter length, you know, and, you know, for a Valentine's Day party. Whereas when you're talking, you're like, I'm going to a Valentine's Day party. I need a dress. It could be this. It could be that. Any of that works, but we just have this multimodal, you know, entry point. And a lot of, you know, people either see a photo, they either see someone they want to photograph or they see stuff on Instagram. And so you can also start with a photo. And then the great thing, really what we've built and spent a lot of time on is understanding the relationship between items in the fashion world.

21:50So a lot of times you'll see something you like, but it's too low cut or it's the wrong color or it's short sleeves and you want long sleeves. And so you can basically pivot what you're looking for based on this combination of image and text. So you can either upload a photo or you can describe what you want. And then when you find something that's similar to it, you can say, yes, I like this, but show me in, you know, green or whatever it is. You can also just, you know, ask for more like a specific item. You can look for something that's higher end or, you know, better priced. And so we've built in a lot of functionality that just is very innate to when you're in the sort of shopping moment and you have these nuances of things you want to be able to express that you haven't had the ability to express in the past.

22:37And so, you know, what we know to be true is that words can express some things. But then once you start showing images, the ability to use the images to help you refine to what you're looking for is quite helpful. And so that's what we've built. And then you get, you know, the search results that are both relevant to the query and relevant to you. We've asked you a bunch of questions up front. And if you skip that, we'll ask you the questions kind of in the context of your search. We store that information and every user gets their own style passport. And so that passport is actually explicit.

23:10There's a place you can go see it. You can update it. You can change it. You can add dimensions to it. But it's sort of the information that you've shared with us and we've gathered from you that helps us predict the things that you'll be likely to want the most as you're going about your different searches. What's really cool about all that, I mean, there's a lot cool about it, but I think the thing that jumped out to me, which resonates most, is this sort of almost like iterative shopping experience. Like, oh, I want, you know, want something green. Oh, not that green, actually this green. And I want the sleeves to be a little bit longer.

23:43Like that sounds really cool. We haven't really seen that before. I imagine that's like one of the key sort of differentiators for why people enjoy this. I'm guessing. I think that's why I will like it. And it's sort of a it's almost like a personal shopper like experience, which, again, reminds me of some of the earlier products you've worked on. Yeah. I think the other benefit is that we bring together all the brands in one place. So most of us, I mean, if you don't, if Google shopping doesn't work for you because the ads that they, you know, are basically the ad section, which is what they're focused on, doesn't get you what you need.

24:15The results are very messy and very inaccurate and very unpersonalized. And so, you know, I think for us, so many shoppers have like 10 tabs open and they're trying to check at this store and this store. They want to just make sure they've seen everything. And so that's the other benefit is we bring sort of all the retailers and the brands together in one place. We're just doing new products. We're not doing secondhand yet. It just is nice to have kind of this comprehensive view that I don't have to go to all these different sites to check. Yeah, it makes total sense. And how do you accomplish this?

24:47Is this through your own models? Are you training your own models? Are you taking, you know, some of the bigger models from OpenAI and Anthropic and fine tuning them? Like, how do you actually technically accomplish this? Yeah, it's interesting because it's actually changed in the last year. the, you know, I think all of us who are sort of working on building applications on top of these LLMs started by really using the LLM directly. So we were working directly with OpenAI. We always test sort of Anthropic and Gemini as well, and sort of just make sure we're paying attention to the latest models and what's doing the best.

25:21And they all have their sort of strengths. Now, what we realized is it's very hard to rely on a large model for a couple of reasons. One is that the answers tend not to be sort of consistent. And so you might get a good answer one time and not a good answer another time. The other is that there's real latency. And so if you are trying to build a consumer-facing experience, the lag time in going to the model, getting the information, retrieving it, and bringing it back is just too long to ask a consumer to wait. And so I think where the world is moving, especially for consumer applications, is this world of ensemble of models, of small models, basically, that learn from, gather information from these big models, and then are much more responsive and targeted at what you need in your zone.

26:11So for us in fashion, we are building mini models on sort of all dimensions of fashion. It just allows us to certainly leverage OpenAI to train models, and then to use that information to build our own mini models that are much more responsive. And over time, what we're doing is we're also adding. So as a catalog of a brand comes into our system, we work directly with brands. We bring in a feed from the brands. And so we're able to kind of augment the information around the product itself. And so we also use AI to understand everything that might be relevant to that product, you know, in the context of how people are shopping.

26:55So occasion-based or body type-based or those kinds of things. Weather-based, you know, this is good for these seasons. So we've built sort of this deep knowledge base around the product catalog and then this deep knowledge base around the user. And then, you know, the models kind of sit underneath the large model to understand the query, quickly serve up the best results for them, and re-rank them based on your sort of personal preferences. And that's kind of how the system works. Very, very cool. Makes a lot of sense. Speaking of the large model providers, I'm sure you've been asked this question 100 times over the past, I don't know, month or so.