How a solo founder used Codex and ChatGPT to launch a fashion brand without engineers | Yana Welinder

Normally I would hire a technical team, like engineers. I didn't, right? I sort of just came to Codex and asked it to build the website. The idea is to use AI really anywhere in the process where that makes sense. It's literally my technical co-founder.

There's so much you couldn't do before, practically from a cost perspective, even from an execution perspective, that is now totally possible. And so you can like unlock your creativity in a way that wasn't possible before.

Really does unlock so much creativity. A lot of my designs start from a hand-drawn sketch. But the other piece of this is I sort of developed the fashion prompt that's behind it, and it describes a lot of sort of what goes into garments. So it has like the, the silhouettes, uh, the proportion or the volume of, of the dress, like kind of like the fabric, how it flows, how it behaves, the construction details, how it moves, even how it sounds. The other piece that is still unsolved, and I'm sort of still banging my head against the wall, is to get it to create patterns. based on my design. And there I'm kind of running humans and AI in parallel. So I am working with human patternmakers, Codex, and just saying, who's going to get to making my thing fastest?

Welcome to How I AI. I'm Claire Vo, product leader and AI obsessive, here on a mission to help you build better with these new tools. Today I have Jana Wellander, and she is building an AI-native fashion startup, and she's going to show us how AI can intersect art, creativity, and real-world production. And why computer use plus software means SaaS really isn't dead, at least not yet. Let's get to it. This episode is brought to you by Merge. Building an AI product is one thing. The hard part is everything around it. Connecting to the tools your team and customers rely on, letting agents take action with the right permissions, and keeping everything reliable and cost efficient once you're in production. Most teams end up piecing that together themselves. So instead of building the products you actually care about, you get pulled into integrations, permissions, routing, and all the infrastructure underneath. Merge is the infrastructure layer for production AI. It connects to thousands of tools, gives agents secure ways to act inside them, and optimizes model routing and spend without you building or owning any of it. OpenAI, Dropbox, and Ramp already use Merge to move fast and build AI right. Visit merge.dev/howiai to start building for free. Jana, thanks for joining How I AI. I'm an art and fashion girl. You know this. There are like 5 of us AI art and fashion girls on Twitter, and we all find each other, but I am one and I love what you're working on. So immerse us in this new thing that you're working on, and then we'll talk about how you're using AI in this really unique fashion.

Cool. Yeah, I'm so excited to be here. Thanks for having me. I am building this AI-native fashion brand, and the idea is to use AI really anywhere in the process where that makes sense. So obviously heavily in the design and production process, but also it's literally my technical co-founder. And so I, that's, that's what I use it for, everything.

And can you show us, like, it's, it's probably hard for people to conceptualize what, like, an AI-powered fashion brand looks like. Can you just like show us a little bit of kind of like some of the stuff that you're working on and then we can get into how it actually, how you actually got there?

So this is Yanabana, the fashion brand, and you can kind of see like every, everything here is built with AI. And so you have some of the different garments that I designed and I'll show later that like different ones start from different, like from different inspiration and from different points. And so oftentimes I start from a sketch. You can see that when you hover over some of these. So I start with, with a hand-drawn, humanly, manually, you know, old school hand-drawn sketch. And then I use AI to turn that into different types of fashion collateral and kind of technical specs and everything along the line to ultimately get to a garment. And I kind of use AI for all the things to visualize it, to create photos, to create videos of some of these garments. There are a lot of them kind of allude to different, like, uh, technical, uh, you know, computer things in a lot of ways, because I feel like that should be really part of the theme. It should be obvious that, uh, that what it comes, comes from. And so the fabrics, uh, will use a lot of kind of the punch card, like computer punch card, uh, patterns and things like that so that you can really see that. And the other The other way in which it sort of appears is you can see, for example, for this like post-keyboard t-shirt, which I'm wearing myself right now, it also is alluding to the fact that now we are not gonna be using keyboards and AI and a lot of this stuff. But can we now have like, give it a second life essentially, and to make sure that it's like obvious the world we live in and the world we're entering?

And I just wanna pause because when I first saw this, I was like, what a fun, like, little art project. Like, I once made this, like, I vibe coded this app, which was, um, Anthropic as an anthropology website, because I thought it was very funny. But you're actually bringing these designs to life. Like, you are actually wearing this keyboard shirt, which is super cool. And so what, you know, I hear a lot of people who are in the creative fields be really fearful of AI because I think it is like displacement in a lot of ways. Like, let's, let's all be honest.

Yeah.

Here, there's like videos and sketching and all sorts of stuff that AI can do. And yet, if you step on the other side of this like post-keyboard future, there's so much you couldn't do before, practically from a cost perspective, even from an execution perspective, that is now totally possible. And so you can like unlock your creativity in a way that wasn't possible before. So I want you to show us step by step. Kind of what goes into creating a garment like this, creating and imagining a design like this, and then actually getting it to production and getting a website like this up.

I think, I think you're absolutely right that it really does unlock so much creativity. A lot of, a lot of my designs start from a hand-drawn sketch, but, um, the other piece of this is I sort of kind of developed this, um, uh, technical, it's essentially a stack, but I've for the purpose of visualizing what it looks like, I have this sort of set up as just like the fashion prompt that's behind it. And it describes a lot of sort of what goes into a garment. So it has like the silhouettes, the proportion or the volume of the dress, like kind of like the fabric, how it flows, how it behaves, the construction details, how it moves, even how it sounds. And I can show in one of the pieces, it kind of has this really interesting kind of sound as you're walking in, in, in the dress. And I take all of this and put it together in essentially kind of when we, if we take it out of my, my, my flow, you can see that it's at the core of it. There's sort of this, this prompt that we could take and plug into ChatGPT. And oftentimes, and when, when I'm working with this, I will also have an actual sketch of what I want and I sort of am then defining this in addition to the sketch. But just for the sake of us kind of playing in real time, I just thought it'd be cool to just, just use prompts to start with. So, so here we have a sharply tailored waist-defined jacket with exaggerated shoulders and blah, blah, blah, right? So now we're putting this in. Oh, I should have told it to make an image, but maybe it'll figure out to do that. Yeah, there we go. Sometimes, sometimes it's smarter than you think, even on instant. And so now it's going to Generate an image. And oftentimes this will give me kind of like a product photo. But usually what I then want to do with this is to then take it to kind of the next level. And I will have very specific types of like collateral that I want to get out of it. So once we have this ready, I will ask it to generate a Vogue-style runway photo. Oh, there we go. So this is kind of like the product photo of this. So it has these exaggerated Shoulders, the waist, a lot of kind of the stuff that I asked it to. And if I had a specific sketch, it would really adhere to my sketch much more so than, than what was in the prompt. But then if we go to here, we say like, now make it into a Vogue style. Can you tell that I normally do not type anymore? And so I usually just use voice for all these things. It's just like so hard.

It's easier for our podcast guests. you use voice because they don't have to imagine what you're typing. So she's, you know, you're typing, make it into a Vogue-style photo.

Of a model walking down the runway in this outfit. And so then I kind of just like iterate on this and there's, I go to photo, that's a runway photo. Then I go to like an influencer photo and you can kind of see this in a lot of, from a lot of different angles. And also iterate on it based on kind of what you're seeing now. Make it, you know, red, make it shorter, make it longer, and all those different things. But I usually just start from, from a specific point and then iterate from there.

I have a question for you, which is why ChatGPT and why ImageGen 2 as the, as the model you're using? There's a lot of ImageGen models out there. Why that particular one?

Yeah. So I've done a lot of experimentation with lots and lots of different image models. And what I found was ultimately Nano Banana and Imagen 2.0, really Imagen 1.5 was already pretty good at this, are the best at following visual directions. And often I'm not doing this. Usually I'm sort of, as I mentioned, I'm mostly giving them a physical, like a sketch. that I want them to turn into a picture. And what Image 1.5 did better than Nanobanana and Image— uh, Images— ChatGPT Images 2.0 does even better is it really well follows the sketch. In fact, it follows the sketch sometimes a little bit too precisely. And so you get something that doesn't look like fabric anymore and you have to kind of be like, no, no, no, no, no, you gotta like, it's gotta flow. Like, this is kind of why I have this. this very detailed kind of prompting where I tell it how the fabric should flow, because otherwise it can get into the mindset of, oh, she really wants it to look like paper, you know. But, but the flip side though, with all the other image models that I've tried, is that they will make it look beautiful and realistic, but it will look very, very different from my design. And obviously when you're designing fashion, you, you need it to look new and different. You don't need it to look like the most similar thing that's walked the runway some other time. And that's usually what you get with a lot of other image models.

What I love, and if you can go back to your prompt flow, what I love about this is this is just, okay, so people are gonna be like, I'm not a fashion designer, I don't have any use for this. What I think is the takeaway here generally for folks is write down your process and write down what a definition of complete or good is. You and I, you know, we're product ladies.

We're product people, yeah, it's a skill.

I think that's what it says. We were like official product ladies and you need a spec, like your prompt is a spec. And so if you can put into the effort of saying like, what makes a really good garment? Like what makes a really good photo? What makes a really good illustration? And come up with like 5 things that you can define for the prompt. You can get a really good output of AI. And then what I love, my friend Zach says, what's good for AI is good for humans. Let's just say you had, you were going to have a human sketch this out or build this. This is the exact information they would need as well to do a good job. And so whether or not you're doing like fashion illustrations or coding, right? Like this, the spec, the PRD matters, people. Like the spec matters. And so getting, forcing yourself to sit down and be a little bit more precise is going to get you that exact outcome that you, that you want. And, um, you know, to be honest, I think that jacket's pretty, pretty rad. I would. I'd wear it. This episode is brought to you by Jira by Atlassian. The teamwork graph in Jira delivers 44% more accurate agent results with 48% less token usage. That's a huge difference when working with AI coding agents like Claude, Cursor, Codex, or Copilot. The hardest part of shipping with AI isn't the code, it's the context. What's the right ticket? What did the spec say? What got decided in Slack? The teamwork graph pulls all of that from across your entire stack, from Jira and Confluence to GitHub, and feeds it directly to your agents before they write a single line. You assign the work, the agent gets everything it needs, and a PR surfaces when it's ready. No digging, no context switching, no broken flow. Same team, smarter agents. Try it free at jira.teamworkgraph.

Awesome.

That's J-I-R-A dot D-E-V. I love it. So you've created this, this image. I also love this idea of taking like a core image and iterating it through like runway photo, influencer photo, like, you know, standalone catalog, like photo product, product shot. What do you do next? Like, what's the next part of your flow?

So a lot of what I do is kind of iterate on it. And so I have this like, you know, change the color, change the, you know, change the fabric. And so I have kind of a lot of these different ways to do that. But then the next piece really comes to kind of like, what is the production process? And the production process is going to be very different depending on what I'm doing. And so to give you An idea from one of these. So for example, I had this, uh, dress that actually did not come from a sketch. It came from— I had this one day where I was just being bombarded with Ruth Asawa's, um, artwork in different places. Like I went to, I went with my son to this like, like kids event and there was an exhibit across the street that I like accidentally, like we were just early, so we went and watched it and we were like at SFMOMA and there's like a permanent exhibit there. Uh, she had a piano recital and, and the church had one of her pieces hanging like literally almost above my head. And I was like, I keep, I don't know. So, and so this was, this is literally from my phone. So I didn't have a sketch, right? Normally I start from a sketch, but in this case I was like, you know, make a photo of a model walking down the runway in a dress inspired by Ruth, Ruth Asawa's Uh, loop-wired sculptures. And it was important to— that it had a beige undergarment because otherwise ChatGPT will not like— it's like, no, you're trying to make me make nude pictures and I don't do that.

Very artistic nude picture.

Just no, I don't do that. Um, actually, I think, I think even here I tried to make it to remove the sculptures in the background and it was like, we're sorry, this violates our nudity guidelines, right? This is like happens all the time.

It's very frustrating. Quick side, you know what, if I showed up to a date with my husband in this gown, I do not think he would be thrilled. I don't think he'd be like, ooh. It's not quite erotic, right?

Exactly. It's not, it's not, it's not nudity. It's not, you know, it's not torn. But, but here we are. And so, so this one, right? Like this is, and I think that's A lot of garments you would do like some, some garments I will drape, right? Like on, on, on the, on the model behind me or some, some garments go into, they become, you create a kind of a pattern. And I have different, some, some AI processes and some non-AI processes that I've done for patterns. But this garment is previously unmanufacturable. And so what I, and I ended up kind of taking this and, and also creating a video of it. This is kind of like, you can see a little bit more of what it looks like. But the next step for this particular garment will be to take it and 3D print it. And particularly these kind of big ball pieces will need to be 3D printed. And so I then went onto Codex and had it help me create CAD files for these balls. So I first prepped it and had it like, I was like, okay, so I made, made that into a sketch. Then I made these pieces into, uh, into kind of more sketch illustrations. Then I, uh, it didn't do well. It still didn't do well. We had a bunch of back and forth. Eventually it started doing something that was a little bit more similar. Um, I had to do, this is a really, really long process. And then ultimately I had it using, uh, computer use, um, go and build this in 3D software so that I could create like an STL file and send, send off, uh, to 3D prints. So that's the process for this particular garment to sort of to, to 3D print it.

I have to, again, like I have to pause because if we're just talking about the creative, like generation process. Again, this is now a garment that would've been previously like almost inconceivable to make, not just because it was, it's difficult to construct, but even if you were like, of course we could 3D print that before AI, we could 3D print it after AI, the like tedium of creating the CAD models was so onerous that like even getting to the point where you could execute on it. Is really hard. And so I just think this moment of unlocking the previously impossible, whether the impossible was technically impossible or if it was just practically infeasible, is super important. The other thing that I love, bless computer use and codecs plus computer use, love you so much because I can just be like, I don't need to know how to use this software. You go use that software. You do whatever you want. I've gone through this like trough of despair of like SaaS is dead. Maybe like we're never gonna touch a website again to, you know what, like agents are really good at pressing buttons. So like software's back for, but agents are gonna use it. It's very, very kind of like interesting shift I'm seeing.

I think that that's a great point. And one, one fascinating piece is that a lot of these things, like for actual end product, like a CAD file, Codex, as amazing as it is, is not great at generating a final CAD file if you asked it to do it on its own.

Right.

But if you use, yeah, if you use computer use and have it go and do it in a piece of software that's designed for that purpose, and I have both for like actually for 3D printing and there's a fashion software called CLO. that I've used as well. And it generates this like, like it generates patterns and it, and it then fits them on a 3D model. And, and I tell Codex to go and use this software that I haven't learned how to use myself. And it does such a great job at something that it itself couldn't do. So it's sort of like SaaS in combination with Codex works so well on a lot of these things that, that don't. And I'm sure it's just a matter of time, but that's, that's where we are today. And I wanna do it today. I don't wanna wait. I don't wanna wait like a year. You know?

I, I love it so much. Okay, so you like generate an idea, we've sketched it, we've created images, we've even started to prototype the construction of it. How are you getting something built? And maybe not the gown, like maybe this like keyboard shirt that you're wearing, like how are we actually getting this manufactured?

Yeah, so for something like that, what I've done, and I do this a lot, is I will go and Now, uh, now this same tool becomes my research assistant and I ask it to identify, like, um, you know, find 10 US-based custom apparel manufacturing companies that are similar to the one that I liked, but I want to see if there's other ones. And then I just like, there's a bunch of things I want and I kick off a deep research, uh, on extra high so that it's extra high while doing it.

Right.

And not in that way. Um, and, um, and then, and then I get back A bunch of things, but usually, so, so for, for a lot of my work, I end up like, I'm doing something else. So I'm like, I am kicking off this research and then I go and like drape fabric on, on model, or I do like, I'm, I'm on my sewing machine doing prototypes, like I'm doing something else. And so a lot of them kind of the next steps end up being via voice. And so I, like in this case, I came back to it and I was like, well, do you see the Um, because I'm talking, uh, apparel and using voice, right? Like the apparel manufacturers that I asked you to find. Um, now can you like put together the emails that I can, that you can send off to them? But like, you may wanna make sure that you get my, my approval to do that. And then I have, and then I use browser, uh, use to let it go into Superhuman and like set all of these up for me. But I really wanna click the Send button before, just to check its work.

Again, like something that is just so tedious. And, you know, if, if I was like, you're, you're, you're much more diligent at this than I am. But if I was like, oh my gosh, I have this vision. We're going to make, we're going to make these keyboard shirts. They're going to be amazing. And then I'd be like, how can I make it? And they're like, ugh, I have to like look up vendors. I would just practically give up because it feels so boring. But if you can, if you can get that work off to somewhere, again, like the next step and the next step and the next step, and then what you're going to end up with is this like incredible business. And so I just am like so inspired by people like you that take AI and create like a niche and like can, can then take that and build out a business where maybe one wouldn't, wouldn't have been available or possible before. Um, I just think it's, it's super cool. Okay. So you're doing this vendor outreach.

Yeah.

You're actually going to get this thing. This thing produced. Side note for all the cool girls out there, um, when they're ready, we all want one. Put me on that. Put me on the preorder, preorder list.

There actually is. I'm gonna, I'm gonna, I'm gonna be that person, but you can preorder it.

Amazing. Perfect. Okay. And then last kind of, kind of flow you have is like, again, going back to like preorders and how to actually run this business, how are you using it to like be your technical co-founder in all this?

That's, I gotta say that that's probably like my favorite piece because this, I mean, apart from the fact that it is like, there's so much AI in this, but it's basically like an AI native, like Everlane or something, right? Like normally I would hire like a technical team, like engineers. I would have possibly have it. I probably wouldn't, let's be real. I wouldn't have a technical co-founder, but like I could have had a technical co-founder. I'm a solo founder girl. Like this is just my nature.

It's the life. It's the life.

It's the life. But I would hire engineers and I didn't, right? Like I had it build this website and I did it and I sort of just came to Codex and asked it to build the website. There's a lot of pieces here that aren't just like a static website in that it, like you can obviously preorder things and I could like add Stripe to it, which is like in prior lives have been like a tedious big exercise, but I can go back and kind of show you the Stripe flow. But adding that was just like, you know, a second. I had to create like databases to be able to like track votes as people are voting on these different garments. So I can decide, kind of figure out which are the most popular ones to bring to life. And so like I had to do that. In, in an earlier phase of this, I had to create a separate dashboard for me to track the votes in real time. And so I did that. Now I really just like, I just use voice and I'm like, oh, hey, how many votes do I have for this thing now? Right? So like, why would I even have a dashboard?

I know.

So I don't do that, but like, just to kind of show you like the Stripe flow as, as, as an example of, of Like, yeah, I came here and I was like, oh yeah, I use browser to add Stripe to my site. It was like, I don't know. You're like, I'm not signed in. I'm like, no, you gotta sign in in my own in-app browser. I'm like, no, no, no, you go to mine. I'm already signed up there. And then I was like, yeah, you gotta do it. I'm like, no, you can do it. You know, we had this like back and forth. It's like, no, no, you can do it. And yeah, this is it. And then I had done. Everything's added and now preorder is live. It's just like a very, very fast process. And literally whenever I like need to change something about like the process, I can really just come here and just like list out all the things I wanna change and it goes off and it finds the result project. It has like, finds the repository on GitHub and does all the things. I'll say like this weekend, my son built his own website using like ChatGPT Sites, and I'm sort of just like sitting there going like, do I, should I really, like, do I wanna migrate off of Vercel? Do I really want, like, do you know, like, is this, is this a whole new, like, now, like, now a 9-year-old can have his technical co-founder like do this and like build. Build a company for him, right? Like, just, it's, it's just so cool to see this all, uh, evolve, you know?

Your, your 9-year-old and my 9-year-old are gonna be best buds. We can create like a, like a Codex club for 4th graders.

We should totally do that.

Um, this is a very, sorry, we are doing San Francisco, uh, Silicon Valley momming stuff right now. I just wanna, I wanna like go back to one thing because ripped from the headlines, Codex prompting. I too am like, hey Codex, do this thing. And it's like, nah, no thanks. I can't. And I'm like, no, I'm pretty sure you can. And then like, after 2, like, I'm pretty sure you can. It's like, you're right. I can. I, um, I'm going to post about this later, but I have this like smart light bulb. You can see it's on my, on my thing. And I was like, hack it. And it was like, no, I really shouldn't. I was like, you can do it. And I was like, done. So just be a little bit insistent.

Yes.

Um, you know, and, and you can get, get stuff done. This has been so, so fun. Again, just like recapping for folks, a business that I think like would probably be inconceivable before AI, fully stacked from the creative process, creating images, creating videos, imagining product ideas all the way through like hardcore manufacturing and sourcing vendors. to just like being able to voice note your virtual CTO and update anything you need to do and have your engineering team on demand in ChatGPT. Pretty incredible. I wanna get through some lightning round questions and then we're gonna get, get you outta here. You know, my first question for you is like, what is the hardest part of this? Like, what is still, maybe like, what is still hard? What isn't solved? What, where are you still feeling friction?

So I'd say it's kind of 2 things. One is that like, Getting it to really do, getting image models to do really unique things per your vision while still making it realistic, that has been really, really hard. I feel like I've solved it, like I've kind of cracked that nut, but it still is. And every now and again, I'm sort of still get something that is really 2-dimensional or whatever, and I have to really refine it. The other piece that is still unsolved, and I'm sort of still banging my head against the wall, is to get it to create patterns. So to get get it to create like actually accurate patterns based on my design. And there I'm kind of running humans and AI in parallel. So I, I am working with human pattern makers and Codex and just saying, who's gonna get to like making my things fastest, like most accurate, like to my vision, you know? And so that's, that's been really cool to see. I'm not, I'm not there yet with either actually, with either.

Okay. I love it. All right. And then my last question that I ask everybody, speaking of like when it's hard and when it fails you and not getting there quite yet, which is how do you prompt AI when it's not working? What's your, are you nice? Are you a nice mom or a mean mom?

I can be both, but I think I'm mostly a nice mom. And the reason is that I find that I get better output. So like, I mean, I am very direct and I'm sort of like, yes, you can. Uh, you know, you go do that. Like, I'm not just like, oh yeah, I'm gonna do this for you. Um, but I'm, I will be like very polite. I will still say please and thank you. Um, and I do think that it's because like the, the, it's, it's trained on body of text. And so if it sees poor text output that is like rude and mean, it will like kind of get in the mind frame of that part of the internet, right? Like versus like. kind of more professional business context, and then it'll deliver something that's a little bit more what I want it to be. So that's, that I, I do it for selfish reasons, not because I'm afraid of the future AI overlords.

Perfect. Well, this has been super, super fun and inspirational and just like a breath of fresh air in terms of a new use case, a creative space that we haven't seen before, and something sort of like manifesting the world. So where can we find you and how can we be helpful to you?

You can find me on X mostly. I am Yana Bana on X. I am now Yana Bana at X. Um, I, I changed my handle to match the brand. Um, but, uh, but yeah, the, in, in terms of just like helpfulness is getting feedback on all this stuff. Like you can come, you can come to Yana Bana and like vote on the garments that you like. Um, so I can get feedback on like what's, what what's working for people and what's not. Like, I'm all about user feedback and customer input. And yeah, tell me, tell me what you like.

I love it. Well, I love Poppy Love. Let's smash that like button. It is very cool.

Awesome. Thank you. I will, I'll let you know when that one is ready to go.

Available for purchase. Perfect. Well, thank you so much for joining How I AI.

Thanks for having me.

Thanks so much for watching. If you enjoyed this show, please like and subscribe here on YouTube or Or even better, leave us a comment with your thoughts. You can also find this podcast on Apple Podcasts, Spotify, or your favorite podcast app. Please consider leaving us a rating and review, which will help others find the show. You can see all our episodes and learn more about the show at howiaipod.com. See you next time.