अनुलिपि 🔗
Jennifer Strong (00:00):
Among the sci-fi futures we've all been promised is that one day our shopping might almost do itself because AI will know our wants, needs, and habits just as well as we do. Leaving aside whether that's something we actually want, it's also a lot harder than it sounds, especially on the retail side where systems are notoriously manual and whether you're a shopper or a store associate, figuring out how to search what you're looking for is a much more complicated task than it might at first seem. Our guest this week is working to build a centralized AI brain for retailers by pulling those fragmented and siloed systems into a single intelligence layer. I'm Jennifer Strong and this is SHIFT. This episode, it's the latest installment of our oral history project.
Nasrin Mostafazadeh (00:50):
So I'm Nasrin, co-founder and CEO of Verneek. We are a deep tech enterprise AI company. We serve predominantly consumer retailers and consumer brands. So what we really build is this deep domain expert AI that becomes their centralized brain that helps their back office employees or their own consumers across the board to really make better and faster decisions throughout their journeys.
Jennifer Strong (01:17):
So you and your husband co-founded Verneek together. You're one of my dear, dear friends. I'm gonna back you up here because you were actually when you started on this, what, like 12 or 14 years old when
Nasrin Mostafazadeh (01:28):
You met? 14, not 12, but yes. Okay. Almost. You know, how many are younger, the two years is like something you fight for? No, I was 14. I wasn't a kid. But no, yes. We were, uh, both. So I was 14, Amit was 15. Um, it's actually a f - really a fun story. Honestly, if, if you would've told me that this would be my life, I would've said you're kidding me that like 20 plus years and I'm still doing the same exact thing. I would've said, "You are absolutely crazy." But it is true. So, um, the, the story goes as follows. Both Amit and myself very independently were interested in the field of robotics and the field of, again, these multi-agent systems that can collaborate autonomously to achieve a certain goal. And so I had an all girls team for, in Iran. Of course, I'm born and raised in, in Iran and this is, you know, back in 2006.
Nasrin Mostafazadeh (02:22):
So I had my all girls robotics team that somehow made it to the international world stage in Germany. Then Omid had his own old boys from a completely different city, by the way, in Iran. Of course, we had no idea that we existed in the universe at this time. So we had both made it to this world stage, luckily. And then little did we know that while competing with all these, like, you know, grad students from MIT, Carnegie Mellon, Stanford, all these American schools that are famous for their robotics and AI teams, we were competing against them. And little did we know that Amit's team would became the first in the world in his own league. Um, and then mine became the second in the world in my own league. There were two different leagues, and I always make this joke that we weren't competing against each other, so there's no telling who is better.
Nasrin Mostafazadeh (03:05):
But anyways, we got introduced to each other very much because it was quite a story, um, back in Iran that everyone wanted me to meet him, everyone wanted him to meet me. And really the rest is, is, is history. And I'm always, whenever Omid and I get tired, we remind each other that we would not have known each other, period, if we weren't the crazy people that we are in terms of our drive and like why do we do all these hard things?
Nasrin Mostafazadeh (04:07):
Funny enough, I don't think I've told you this. I hope I haven't. So the real story of how we switched from robotics to what we do literally to this day is when we came back from Germany, the reason we could win was that, uh, it was a result of a year or more than a year of having skipped classes, right? So we wouldn't go to our normal day-to-day math or, uh, science or whatever, like, uh, classes. So then we came back, of course, the understanding was that, that they would allow, like, we are now, you know, national, uh, achievements, that they should allow us to go to the next grade, although we don't show up in classes. But then there was one teacher that said no. Oh. And that teacher, yes. And that teacher, believe it or not, was a phys ed class teacher. And she was like, "Isn't that funny?" Uh, and then so she, she sits me down.
Nasrin Mostafazadeh (05:00):
She's like, "There's no way I would allow you to graduate to go to the next grade unless..." She hands me th- these many papers. She says, "Unless you translate these papers from English to Persian."
Jennifer Strong (05:13):
And for those who can't see, Nasrina is showing us a very large stack, like many inches tall.
Nasrin Mostafazadeh (05:19):
Yes, yes. Many inches tall of, like, literal volume. And then, you know, like, that day, I, I go home and I'm like, "My gosh, there's no way I can do this." But you know what? I'm sitting in front of my computer every single day. We are trying to automate a lot of different tasks, again, with our AI. Shouldn't there be a way to automate this? Shouldn't there be a way to automatically translate, you know, from a language to another? And of course, lo and behold, I Google, like, how to translate from English to Persian and, you know - . Whatever, vice versa. And the first blue link, back in time, blue links were a thing, kids. So you would come to the first blue link, which was Wikipedia, um, and of course the, the, the, the Wikipedia page was machine translation. So I click, and I read and read and read, and it explains how machine translation is a subfield of natural language processing.
Nasrin Mostafazadeh (06:14):
And that natural language processing is, of course, a subfield of artificial intelligence, which was a field I was already working on. And that
Load full transcript
was literally it. That is how I got into the field of natural language processing, natural language understanding. The part of this story that you may remember is that, um, I always tell the story of that monkey ate the banana when it was hungry and how you try to decipher basically the right pronoun resolution, uh, being a, the reason I got into common sense reasoning, which is even deeper subfield of, you know, natural language processing. And the reason I even actually got into that was when I kept reading on Wikipedia, I got to this page that was explaining, um, the problem of, at the time, you know, they were calling it AI complete problem. So in computer science, as you know very well, we have MP complete problems, which are these problems that are very, very hard to basically solve.
Nasrin Mostafazadeh (07:06):
And if you solve them, you would solve the rest of, you know, computer science problems. So there was a thing called AI complete, meaning that if only you solve for this class of, uh, problems in AI, you would solve everything for AI. And that being common sense reasoning and solving that monkey ate the banana is something I read on Wikipedia. And that's why to this day I still, if I can, I donate every dime I have to Wikipedia. And I'm so grateful that they really, truly, uh, democratize access to knowledge that way, that a kid like me growing up in Iran could really get exposed to that level of information so quickly and so easily that it changed my whole life, right, for the better.
Jennifer Strong (08:12):
Going back to Verneek, you've been pretty explicit that generic LLMs are not enough for retail. So what do they get wrong? What does domain-specific AI solve that a chatbot doesn't for this field?
Nasrin Mostafazadeh (08:23):
Mm, that's a very, very good question. And honestly, I don't think it's just about retail, but it's more general about enterprises. I think if you think critically enough, pretty much anything that has solved for the bottom line of a given enterprise has never been generic unless it was a generic function, right? And so back to retail, retail is filled with domain-specific problems, right? The problem of, uh, merchandising, the problem of planning, the problem of allocation on the back office, right? The problem of pricing, right? The problem of, uh, making it so that a CEO can actually understand the trends and make, again, more holistic end-to-end bets on, uh, for their better decision making, and then all of it on consumer facing the pro - you know, problem of styling, um, you know, what outfit to, to put together, the problem of, you know, finding the exact right item through searching to then trickle down to not just adding to basket, but actually closing that purchase.
Nasrin Mostafazadeh (09:21):
Uh, so these are all very, very specific problems that need domain specificity. Of course, the dream of the dreams is one day, uh, those that, you know, pursue AGI would believe that the, these, you know,
so-called, uh, artificially general intelligent being can do it all, but that, until whether or not, you know, that, uh, as a goal exists, which I, as a scientist, I would argue AGI itself is a myth, but whether or not that eventuality will come as it stands, the AIs that we have are not deep domain experts in all these little functions, but what we have as an engine at Verneek is the capability of creating that deep domain expert for all these different functions of retail in a very domain specific, very retail specific way, which is why what we do in the market really translates into immediate business ROIs and really a long-term, uh, gains that the, you know, retailers specifically, as well as consumer brands, are in desperate need of.
Jennifer Strong (10:24):
Nordstrom is, I would say, maybe one of your anchor clients. Correct me if I get something wrong here. What changed operationally once your system was in place there?
Nasrin Mostafazadeh (10:32):
So when you go, for instance, as Jennifer shop for an item, let's talk about nordstrom.com as a website, right? When you click on a product page, it has lots of details on it, right? It maybe says that this, you know, particular dress that I'm wearing has this shape, this color, this pattern, right? Um, uh, and like this high, you know, length of the, the sleeves and so on and so forth. And all these attributes are the ones that determine how the, uh, you know, a given retailer actually merchandises this item, right? Because they may have forecast. I'll give you an example. They may have forecast that the color, I don't know, pink is gonna be trendy this year, and now there are a bunch, you know, thousands and thousands of merchants at Nordstrom or any retailer whose job is to go and basically merchandise anything, uh, pink that they can get their hands on within a certain, you know, budget.
Nasrin Mostafazadeh (11:24):
So that becomes a task on the back office that requires knowing not just the color, but shape, et cetera, et cetera, of a given item. So this process of knowing what the attributes, uh, and really descriptions of your products are has been historically done very manually, which blows my own mind, but that's how the entire world of retail predominantly runs even to this day when we are not part of the equation. So case in point at Nordstrom, they would be having a lot of manual labor that was, you know, hundreds of people whose day-to-day job was to manually attribute tag the items so that they could be, uh, merchandise, like bought, right? And then ultimately you have to just make sure how many to buy and, you know, where you warehouse it and then where to allocate it, which store to ship it to, et cetera.
Nasrin Mostafazadeh (12:12):
And then ultimately, how to put it on your website, right? Your digital, basically the pr - you, you know how when you go to a website and you're trying to find some information about that item, a lot of that historically has been generated, uh, manually, uh, although the vendors, quote unquote, being the brands that sell this, you know, whatever dress to Nordstrom should give them that information, but every retailer has their own specification. So anyways, there are a lot of hands involved and a lot of bespoke needs as to what it takes to just set up a product so that you and I can go and shop it. And so that process for Nordstrom, which again, it takes, this is change management, so it takes years and years and years to see the ul- ultimate outcome. They ended up saving about 90%, right, in terms of the efficiency gains.
Nasrin Mostafazadeh (12:57):
They ended up going about 20 times faster in just the time that it takes to set up an item so that they can sell it, right? Because when it's manual, not only it's, you know, it just takes a very, very long time with all the humans that are involved. And then ultimately, talking about accuracy, we showed that our AIs in, that can, you know, see and touch and feel the stress, uh, using the image of it and all the little bits and pieces of information that we can triangulate from, is 20% more accurate than the most caffeinated human being, of course, whose job has been to do this, who didn't even wanna do that job. It was on top of million other actual strategic things that they were hired to do, but they just had no choice but to do this as well because as a merchant, they couldn't have bought it.
Nasrin Mostafazadeh (14:09):
In the beginning of the year, there were all these LLM companies, the chat bots, that wanted to get into the game of commerce. You may recall the, the Shopify integration for ChatGPT was launched, the, even some of the retail integration with likes of Walmart and ChatGPT had launched. And the idea was that what if this chat interface can also help you shop, can also help you get, you know, your general information, can also help you code, et cetera, et cetera, which of course, to a good degree, can. But what has happened now in the hindsight, um, you know, we are in September 2026 is that a, a majority of those aspirations have failed so much so that, you know, we are at the cutting edge of seeing, right, what is the, uh, incoming traffic from these, you know, LLMs to our retail partners, to their websites, to their stores, et cetera, however one can track.
Nasrin Mostafazadeh (15:06):
And none of it has been actually major. And why is that? It's because the way that the integrations happen for generic. And I think the right way requires that, again, deep domain understanding of what shopping experiences that people like are. And I'll give you some, uh, examples. Like what we've learned very, very early on in our journey when we were in our very first physical stores that we had launched our AIs, we saw firsthand that, for instance, when you have a shopper in the aisles, no one wants to read a paragraph after paragraph after paragraph of where can I find, you know, the healthiest peanut butter that doesn't have sugar in it, right? So even the most basic design decision of what is the length of the response has everything to do with how efficacious that interaction can be, let alone how you should actually visually make it compelling.
Nasrin Mostafazadeh (15:58):
And now that we work predominantly, you know, in fashion and luxury, we see that shopping is very visual, right? And believe it or not, actually, the least of it is the text that you generate. The least of it is the text that can describe what, you know, is basically the right answer. And a lot of it is actually an intelligence that is very, very multimodal, right? We automatically curate, like, looks, we put, like, little building blocks together that tell a story, right, in a very visual way. And there's a reason, again, these experiences have been historically what, you, you know, a lot of shoppers have gravitated towards, and there's a right way of making it in an AI native way, and there is a bolted on, again, way of making it, um, uh, basically go in the hands of the shoppers. So to perhaps sum it up, my projection was, this is the year of AI for shopping, which rightfully so, we've seen all the activities in trying to get in because of the, all the possibilities and all the impact that I was mentioning it will have and has in terms of its reach and the context and so on and so forth.
Nasrin Mostafazadeh (17:04):
But of course, it's way, way harder than any of these companies had anticipated. And, um, we do see now that more and more retailers and consumer brands are interested in, again, that these more unbrand, deep domain expert ways of going about adding AI as a part of their user experiences.