Julie Bornstein concept it could be a cinch to put into effect her thought for an AI startup. Her résumé in virtual trade is impeccable: VP of ecommerce at Nordstrom, COO of the startup Sew Repair, and founding father of a customized buying groceries platform obtained by way of Pinterest. Model has been her obsession since she used to be a Syracuse top schooler breathing in spreads in Seventeen and putting out in native department stores. So she felt well-positioned to create an organization for purchasers to find the very best clothes the usage of AI.
The truth used to be a lot more difficult than she anticipated. I had breakfast lately with Bornstein and her CTO, Maria Belousova, to be told about her startup, Daydream, funded with $50 million from VCs like Google Ventures. The dialog took an sudden flip as the ladies schooled me at the unexpected problem of translating the magic of AI techniques into one thing other folks in truth to find helpful.T
Her tale is helping provide an explanation for one thing. My first e-newsletter of 2025 introduced that it could be The 12 months of the AI App. Even though there are certainly many such apps, they haven’t reworked the arena as I expected. Ever since ChatGPT introduced in overdue 2022, other folks were blown away by way of the tips carried out by way of AI, however find out about after find out about has proven that the generation has now not but delivered an important spice up in productiveness. (One exception: coding.) A find out about revealed in August discovered that 19 out of 20 AI endeavor pilot tasks delivered no measurable price. I do assume that productiveness spice up is at the horizon, but it surely’s taking longer than other folks anticipated. Paying attention to the tales of startups like Daydream which are pushing to wreck via offers some hope that endurance and endurance would possibly certainly make the ones breakthroughs occur.
Fashionista Fail
Bornstein’s authentic pitch to VCs appeared glaring: Use AI to unravel difficult style issues by way of matching shoppers with the very best clothes, which they’d be overjoyed to pay for. (Daydream would take a minimize.) You’d assume the setup could be easy—simply hook up with an API for a type like ChatGPT and also you’re excellent to head, proper? Um, no. Signing up over 265 companions, with get admission to to greater than 2 million merchandise from boutique retail outlets to retail giants, used to be the straightforward phase. It seems that pleasing even a easy request like “I want a get dressed for a marriage in Paris” is amazingly advanced. Are you the bride, the sweetheart’s mother, or a visitor? What season is it? How formal a marriage? What remark do you need to make? Even if the ones questions are resolved, other AI fashions have other perspectives on such issues. “What we discovered used to be, on account of the loss of consistency and reliability of the type—and the hallucinations—occasionally the type would drop one or two components of the queries,” says Bornstein. A person in Daydream’s long-extended beta check would say one thing like, “I’m a rectangle, however I want a get dressed to make me appear to be an hourglass.” The type would reply by way of appearing attire with geometric patterns.
In the end, Bornstein understood that she needed to do two issues: delay the app’s deliberate fall 2024 release (despite the fact that it’s now to be had, Daydream continues to be technically in beta till someday in 2026) and improve her technical group. In December 2024 she employed Belousova, the previous CTO of Grubhub, who in flip introduced in a group of best engineers. Daydream’s secret weapon within the fierce ability struggle is the risk to paintings on a captivating downside. “Model is one of these juicy area as it has style and personalization and visible knowledge,” says Belousova. “It’s a fascinating downside that hasn’t been solved.”
What’s extra, Daydream has to unravel this downside two times—first by way of decoding what the client says after which by way of matching their occasionally quirky standards with the wares at the catalog facet. With inputs like I want a revenge get dressed for a bat mitzvah the place my ex is attending together with his new spouse, that figuring out is significant. “We’ve got this perception at Daydream of purchaser vocabulary and a service provider vocabulary, proper?” says Bornstein. “Traders discuss in classes and attributes, and consumers say such things as, ‘I’m going to this tournament, it’s going to be at the rooftop, and I will be with my boyfriend.’ How do you in truth merge those two vocabularies into one thing at run time? And occasionally it takes a number of iterations in a dialog.” Daydream discovered that language isn’t sufficient. “We’re the usage of visible fashions, so we in truth perceive the goods in a a lot more nuanced manner,” she says. A buyer would possibly proportion a selected colour or display a necklace that they’ll be dressed in.
Bornstein says Daydream’s next rehaul has produced higher effects. (Even though after I attempted it out, a request for black tuxedo pants confirmed me beige athletic-fit trousers along with what I requested for. Hiya, it’s a beta.) “We ended up deciding to transport from a unmarried name to an ensemble of many fashions,” says Bornstein. “Every one makes a specialised name. We’ve got one for colour, one for material, one for season, one for location.” As an example, Daydream has discovered that for its functions, OpenAI fashions are truly excellent at figuring out the arena from the clothes standpoint. Google’s Gemini is much less so, however it’s rapid and exact.
