Anabel Maldonado

The through-line

Anabel Maldonado's whole argument rests on one observation from her years inside luxury e-commerce: the industry had no explanation for the most important thing that happens to a shopper. She describes the moment of seeing a pair of shoes or a dress "where we're like, that's me. Or that's absolutely not me" as "such a deep, visceral reaction," and says the striking thing was that "no one was really looking into that" [3]. Everything she has built since follows from treating that reaction as measurable rather than mysterious. Her frustration with the alternative is consistent and specific: "I hated how we were always, how we were promoting things, you know, it was always florals for spring" and "here's bestsellers the editors are wearing" [3][5], a mode of merchandising with "zero inquiry into you know why an individual likes what they like" [5].

The second, equally persistent theme is that she sits between two camps that do not talk to each other. "Part of the reason why this technology doesn't exist yet is the quantitative and the technical sits very far from the creative and the qualitative," she says, "and they don't know how to speak to each other" [3]. She frames PSYKHE AI as "the bridge between the creative and the analytical" [4], and attributes the persistent weakness of retail recommendations to the fact that "that understanding is lacking by the people who are building it" because "they don't understand the problem in an analog sense" [5].

On why taste has a structure, and the Big Five is it

Maldonado's framework starts from the Big Five, or OCEAN: openness, conscientiousness, extroversion, agreeableness and neuroticism [3][5]. She defends the model on scientific grounds rather than convenience: "the reliability and the validity of this model is very strong," it has been studied "across different populations geographically," and "your big five traits you're born with them you can see them in infants and they're pretty stable" [5]. She walks through each dimension in ordinary terms, from the high-openness shopper who wants "the latest trends" and things "a little bit out there" against the low-openness traditionalist who likes routine, to high conscientiousness as an obsession with "status and achievement" versus the disorganised free spirit, to agreeableness as people-pleasing against "the classic contrarian," to neuroticism as proneness to negative emotion paired with creativity, "that tortured artist stereotype" [5]. The traits are spectrums, not boxes: "you can fall from like one to 100 technically" [5], and the combinations run to "over 3000," "which is why it's a good AI problem. It's not something a human's really gonna be able to put together" [3].

What makes the model useful commercially is that it correlates with things far outside psychology. She cites health and divorce statistics, political leanings, food and music preferences, notes the traits were used to influence the 2016 election, and uses herself as the worked example: her hatred of florals and pastels traces to being low in agreeableness, so "I need fashion that's more severe" [5]. The same trait, she points out, predicts a preference for bitter flavours, whiskey and black coffee over sweet drinks, and it shows up in travel too, since she avoids "mass destinations where everyone agrees that they're good" in favour of things off the beaten path [5]. This is why she insists the domain is taste rather than fashion: the same sensibility governs "a hotel room that I'm staying in or architecture in a city that I'm traveling to or a tablescape at a dinner party" [5], and elsewhere "a travel destination a hotel suite A Car a hoodie a couch" [4]. She has also explored the reverse direction of the relationship, how clothes affect our behaviour and how we think about ourselves, and whether fashion and style are different things [12].

On why conventional recommendation engines fail in aesthetic categories

Her critique of the incumbent approach is that it was built for constrained, repeat-purchase data sets. "If you watch a Liam Niss movie, it's very high probability you're going to want enjoy another action movie. If I listen to late '9s hiphop, I'm going to listen to late '90s hip-hop again. But if I bought an oversized taupe blazer, I don't need another oversized taupe blazer" [5]. The same logic applies to grocery and CPG: "I'm not buying the same cat food over and over again," so models that work "okay" for functional purchases "don't work in in a lot of consumer products that are ... aesthetically driven where it's emotional" [5]. She makes the point again with apparel: "you have this black vest, you've got one, you don't necessarily need three other black vests, but if we get the ... the quality of what you like about it, we could recommend so many other things to you" [3].

She calls the carousel-and-visual-similarity approach "personalization 1.0" [2]: "a very um you know simple cap boost model only in the carousels," resting on "this assumed intent to buy something specific ... such as you know you've looked at a blue t-shirt Show You Blue t-shirts," a "very rope mechanical expression" that she credits with paving the way but regards as a ceiling [2]. She traces it to an industry habit of asking "what's good enough what's a scrappy scalable way to create a bit of personalization" [2]. Models trained purely on intent inherit the same limit: if it is "optimized for conversion and clicks and I have been looking at a black hoodie ... that's its best guess you kind of can't blame it," but "there's no evidence that she definitely needs to buy a black vest" [4]. Her benchmark for a better outcome is human: "a really good salesperson that knows you would actually like surprise and Delight you with something that's an adjacency ... like oh this came in and I thought of you" [4]. She is blunt that the current state of the art has not earned consumer affection: "you've never heard of a consumer really saying wow this really you know the store really gets me this personalization really works" [4]. When it does work, she says, the shopper notices nothing except "wow, there's so much stuff I like here" [3].

On discovery, not search

A related conviction is that the industry systematically overestimates how purposeful shoppers are. Industry professionals take their own acuity for granted, knowing before a trip exactly which brands, sizes, colours and lengths they want and using the filters accordingly, "but what's interesting is most shoppers don't shop that way," they "don't know what they want with that kind of acuity," and the job is to unburden them [3]. She extends the criticism to the newer wave of agentic demos built around a "customer looking for a dress to wear to a wedding in Costa Rica": "more often it's the other times around you just see something you really like you get that hit of dopamine you get that gut level check and then you find a place to wear it" [5]. Her conclusion is operational: "the majority of customers in these aesthetic verticals aren't searching. You need to get the thing in front of them" [5].

On reranking the entire catalog

The product expression of all this is scope. Existing tools, she argues, leave most of a catalog unreachable, and she is explicit about the arithmetic. Some retailers she works with have 18,000 products in one category, roughly 400 pages, while "the average consumer clicks through 2.5 pages and then we're done" [3]. If a category is static or at best floats a few best sellers to twenty million people, "the thing that might convert URI might be on page you know 212 and we'd never see it it just cannot convert because it's never been seen" [2]. Hence the claim that separates PSYKHE AI from search vendors and carousel vendors alike: "we're taking that, you know, you go on Nordstrom, there's 110,000 products in all men's. We're ranking that for every single user. Every single user sees a completely different page," bringing "that dynamism of Tik Tok on Instagram" to category pages [5]. The tool reranks the selection in real time as the user interacts, roughly every ten seconds [9]. She frames the whole thing by analogy: "if you think about how big tech works, Instagram, Spotify ... they're just large recommendation engines," so two people following identical accounts still see a different order [3][4].

Technically she describes three components. First, product intelligence, using the retailer's product information and LLMs to build "very specific embeddings," which she explains as "a long string of numbers that ... describe where that product sits in the in the vector space ... like coordinates on a map," calling these embeddings the "secret sauce" that lets the system "find relationships between disperate items very quickly" and probably "the most spec you know multi-dimensional ones that exist today" [5]. Second, tracking everything users do, training "individually" and "collectively" [5]. Third, a deep learning model that predicts what will lead to a purchase, which is what ranking is: "for Michael, this is the most likely, this is an X most likely" [5]. The personality layer sits inside the feature engineering, with the model assigning psychological scores to products so it understands its natural consumer: "A broke couch gets a score, um, an oversized logo hoodie gets a score" [3]. She stresses this is a numeric version of language the industry already uses when it calls something "romantic or directional or avante guard" [3], and that the personality model runs "along with five other non personality type models, neural networks" [3]. She is careful not to overstate the psychographic layer: "we're not saying one thing is more important than the other we're just feeding it all the information it needs ... sometimes you're just looking for the cheapest pair of Air Force Ones in a in a size nine and we can pull those up too" [1]. On results, she reports conversion rate increases "between 3 and 5x" and engagement metric increases of "up to 168%" [1].

On who the technology is for

She defines the ideal customer by a single word: "the key word is really variance. Like the more variance there is in the products, the more there is a discovery problem" [5]. That points to multibrand, multicategory retailers as the core market [3][5], with department stores and marketplaces as the natural affinity as the company grows [5]. She sets concrete floors rather than ceilings: "no retailer is too big," but she tends not to work with retailers under 50k uniques a month or with less than four pages of product in their largest category, and typically starts with retailers around $70 to $75 million and up [5]. The stakes she attaches to the work are inventory-level as much as customer-level: when product is never seen, "you don't know why. It's, why it didn't sell. Did the right consumer actually even see it" [3].

On what AI changes for retail more broadly

Beyond her own product, she expects AI to "automate a lot of things that should be automated" and, "more than anything," to make the industry "a lot more intelligent" across every area [3]. She singles out two applications outside aesthetics: personalising on sizes and price points so the right things reach the right consumer, and feeding demand intelligence back upstream "to inform buying and creative decisions" [3].

On competing with big tech, and on founder-market fit

Asked about the scale advantage of Amazon, Google and Microsoft, her answer is that scale is not the binding constraint: "they have access to obviously lots of data points, but you know, data alone isn't enough. You have to be able to, to build it the right way and ask the right questions and have some particular understanding" [3]. The defensibility she names is a patent covering the part of the system that assigns psychographic scores to products, which she describes as protecting the ability to use psychographic data and personality traits in that way [3][4], plus "specialist knowledge" and "expert know-how bred out of my experience" [3]. She invokes Peter Thiel's argument that "great companies are built upon secrets secrets about nature and secrets about people" as the reason incumbents have not solved this [5]. Her account of her own path is the same argument in personal form: she never planned to run a technology company, "I always just followed my curiosity in my, my, my obsession" [3], and she insists the differentiation is downstream of that authenticity, since "there was never a point where I'm like well you know I want to be a tech entrepreneur and make lots of money" [4].

On fundraising and managing your own psychology

Her advice on investor meetings starts somewhere most founders do not. "The most important thing is really about kind of preparing your energy," because what investors pick up on most is "does this person believe what they're saying do they have conviction in themselves does this feel inevitable" [4]. She treats the decision as emotional first and logical second: they then check whether "everything she say actually check out and make sense logically," but "the former is much more powerful" [4]. Citing Ben Horowitz, she says "the hardest thing about being a Founder is managing your own psychology," and that keeping your energy right "takes a lot of effort" when everything around you is stressful [4]. The practical corollary is scheduling: do not book investor meetings back to back, especially the key ones, because a bad meeting caused by someone else's off day will bleed into the next, and you need room to recover [4]. Her own routine is unglamorous, a few hip-hop songs and a sweater, on the principle of "do what you need to do to feel good ahead of that meeting" [4]. She pairs that with having the numbers written down beside you, what is in the bank, how much is left to raise, ARR, "because you don't want to be um you know umming and aing" [4].

She is equally clear that a first meeting, for her about half an hour, is a two-way assessment happening simultaneously, not sequentially [4]. She rarely feels the need to ask set questions because "I've picked up on tons of little cues" from tone, from what an investor chooses to talk about, from how much they listen [4], and she has a hard rule of thumb: "I've never had like a really quiet meeting that all of a sudden turned out well" [4]. She sorts outcomes into three buckets: the immediate yes, the investor who gets it but wants to see trajectory and meet the team, who "usually" says yes later if you stay consistent and show them the right things, and the ones with no real connection, where following up too much is wasted energy [4]. She rejects closing techniques outright: "no one wants to feel like they're talking to you know like a Salesman," and while a better technique might have won her one investor, "for one of those there's probably three or four that are really like I really trust her really quickly because I can sense that she's being genuine" [4]. The antidote to desperate energy is detachment from the outcome, an idea she borrows from dating, believing "your life and your happiness doesn't depend on this one call" [4]. She has raised over a million in total, half a million of it with advisory support [4].

Takeaways

• Repeat-purchase and visual-similarity models break down in aesthetic categories: "if I bought an oversized taupe blazer, I don't need another oversized taupe blazer" [5], and a black vest owner does not need three more [3].
• The core commercial problem is unseen inventory: with 18,000 products in a category and an average shopper clicking 2.5 pages, a converting product on "page 212" never gets a chance [2][3].
• PSYKHE AI reranks whole categories per user in real time rather than optimising carousels, reordering 110,000 products in a retailer's men's section so every user sees a different page [5], with reranking every ten seconds as the user interacts [9].
• Reported outcomes are 3 to 5x conversion rate increases and up to 168% increases in engagement metrics [1].
• Personality is one input among many, not a substitute for basic intent: "sometimes you're just looking for the cheapest pair of Air Force Ones in a in a size nine and we can pull those up too" [1].
• Fit is defined by variance: multibrand, multicategory retailers, generally 50k+ uniques a month, four or more pages in the largest category, and typically $70 to $75 million in revenue and up [5].
• Against big tech, her position is that "data alone isn't enough" [3], with defensibility resting on a patent for assigning psychographic scores to products [3][4].
• For investor meetings, prepare your energy before your research, space key meetings apart so a bad one does not contaminate the next, keep your key numbers on a notepad, and use no closing techniques at all [4].