Research
PSYKHE AI began with a simple hypothesis: taste isn’t random - it’s patterned, and those patterns map back to among other things, stable psychological traits. Personality is an important predicitive layer.
It started with a SurveyMonkey study of 327 respondents where we explored correlations between aesthetic preferences and Big Five personality dimensions. We chose the Big 5 because it's proven to be the most predictive and scientific personality model in psychology. The signal was clear: there were statistically significant correlations between aesthetic choices and underlying personality traits. That early data became the seed for our larger scientific initiatives.
We then ran a large-scale study using the official shortened BFI-2 (copyright 2015), with permission by Oliver P. John and Christopher J. Soto. Over 100,000 consumers completed validated psychometric assessments and interacted with a diverse set of fashion and interior products. Their feedback was highly granular. People could “zap” items they disliked, save, purchase, and express strong directional signals across more than 3 million SKUs spanning several product categories.
This gave us an unusually rich alignment dataset: explicit trait scores paired with fine-grained product interactions. It allowed us to build mappings from personality structure → aesthetic preference → product attributes - a relationship that has held remarkably stable across domains.
Today, that foundational research powers our ingestion layer. Using computer vision, LLMs, and NLP, we embed and project new products directly into our psychographic space, giving each item a profile that mirrors how real people with real traits respond to it.
We’ve continued strengthening and validating this psychographic framework with multiple datasets and consumer segments, expanding its robustness while keeping the core methodology proprietary.
It started with a SurveyMonkey study of 327 respondents where we explored correlations between aesthetic preferences and Big Five personality dimensions. We chose the Big 5 because it's proven to be the most predictive and scientific personality model in psychology. The signal was clear: there were statistically significant correlations between aesthetic choices and underlying personality traits. That early data became the seed for our larger scientific initiatives.
We then ran a large-scale study using the official shortened BFI-2 (copyright 2015), with permission by Oliver P. John and Christopher J. Soto. Over 100,000 consumers completed validated psychometric assessments and interacted with a diverse set of fashion and interior products. Their feedback was highly granular. People could “zap” items they disliked, save, purchase, and express strong directional signals across more than 3 million SKUs spanning several product categories.
This gave us an unusually rich alignment dataset: explicit trait scores paired with fine-grained product interactions. It allowed us to build mappings from personality structure → aesthetic preference → product attributes - a relationship that has held remarkably stable across domains.
Today, that foundational research powers our ingestion layer. Using computer vision, LLMs, and NLP, we embed and project new products directly into our psychographic space, giving each item a profile that mirrors how real people with real traits respond to it.
We’ve continued strengthening and validating this psychographic framework with multiple datasets and consumer segments, expanding its robustness while keeping the core methodology proprietary.
JEPA Embeddings
We didn't stop at psychographics. We wanted the full picture of all the factors that go into a taste-driven decision.
At the core of PSYKHE AI’s system is a JEPA-style embedding architecture - models that don’t waste compute predicting pixels or reconstructing images, but instead learn the latent structure behind human taste.
Traditional models chase surface correlations. JEPA models do the opposite: they learn stable representations by predicting the relationships between things, not the raw things themselves. We learn why a shoe, hotel, or car fits their deeper pattern of preferences.
Our architecture uses a hybrid of language and vision backbones trained to encode products and user interactions into a shared latent space. Because the system learns structure - not noise - it generalizes across categories (fashion, interiors, CPG, specialty retail (i.e. toys), grocery, travel) without retraining from scratch.
The result: a persistent, psychographically meaningful embedding for every product and every user, updated in real time as new context arrives.
At the core of PSYKHE AI’s system is a JEPA-style embedding architecture - models that don’t waste compute predicting pixels or reconstructing images, but instead learn the latent structure behind human taste.
Traditional models chase surface correlations. JEPA models do the opposite: they learn stable representations by predicting the relationships between things, not the raw things themselves. We learn why a shoe, hotel, or car fits their deeper pattern of preferences.
Our architecture uses a hybrid of language and vision backbones trained to encode products and user interactions into a shared latent space. Because the system learns structure - not noise - it generalizes across categories (fashion, interiors, CPG, specialty retail (i.e. toys), grocery, travel) without retraining from scratch.
The result: a persistent, psychographically meaningful embedding for every product and every user, updated in real time as new context arrives.
What’s Patented
PSYKHE AI holds an issued U.S. patent for a system that assigns psychological trait scores to products and matches them to users based on underlying psychographic compatibility. The patent covers how we transform product data into a structured “personality profile” and place users and items in the same vector space to enable deeper, trait-driven recommendations. It has been cited by companies such as Microsoft and Walmart, underscoring the technical maturity and forward-looking nature of this feature-engineering methodology.
Measurement & Validation
We validate PSYKHE AI the same way we build it: with rigor, structure, and real-world data.
Every deployment undergoes causal A/B testing to isolate the lift driven by our psychographic world-model from surface-level noise. We measure impact using retail-grade metrics - RPV, conversion rate, CTR, AOV, and downstream engagement - tracked in real time across the full customer journey.
Our JEPA-style embeddings consistently outperform collaborative and popularity-based baselines by capturing latent preference structure, not reactive co-click patterns. Because each user’s taste vector is persistent and context-aware, performance remains strong even in low-data environments, reducing cold-start problems that plague traditional recommenders.
Across pilots and post-pilot rollouts, retailers see 10–30% sustained Revenue-Per-Visitor (RPV) uplift, validated through strict statistical standards to ensure the gains are causal, repeatable, and attributable to the architecture itself. This is how we demonstrate that PSYKHE AI functions as an intelligence layer.
Every deployment undergoes causal A/B testing to isolate the lift driven by our psychographic world-model from surface-level noise. We measure impact using retail-grade metrics - RPV, conversion rate, CTR, AOV, and downstream engagement - tracked in real time across the full customer journey.
Our JEPA-style embeddings consistently outperform collaborative and popularity-based baselines by capturing latent preference structure, not reactive co-click patterns. Because each user’s taste vector is persistent and context-aware, performance remains strong even in low-data environments, reducing cold-start problems that plague traditional recommenders.
Across pilots and post-pilot rollouts, retailers see 10–30% sustained Revenue-Per-Visitor (RPV) uplift, validated through strict statistical standards to ensure the gains are causal, repeatable, and attributable to the architecture itself. This is how we demonstrate that PSYKHE AI functions as an intelligence layer.
Privacy & Security
PSYKHE AI is built to meet modern data standards from the ground up. We use fully anonymized user IDs and do not require any PII to generate or update psychographic embeddings. Our systems are engineered to comply with SOC-2, GDPR, and CCPA, with strict controls around data access, storage, and retention.
All processing occurs within secure cloud environments, and our models operate on behavioral and product-level event streams rather than personal identifiers. This architecture keeps user privacy intact while enabling retailers to deploy a high-precision intelligence layer without introducing additional compliance burden.
All processing occurs within secure cloud environments, and our models operate on behavioral and product-level event streams rather than personal identifiers. This architecture keeps user privacy intact while enabling retailers to deploy a high-precision intelligence layer without introducing additional compliance burden.
Founder’s Note
"I believe that when our external world aligns with our internal values, life takes on a kind of magic. We crack the code of consumerism when we understand that consumption is rooted in looking for psychological alignment. After function and price is satisfied, why do we choose A over B?
Everything we’re drawn to - cities, homes, art, cars, music, travel destinations - isn’t just about surface features, but about the deeper qualities an object or place embodies. We seek that which reflects who we are, and who we aspire to be. If we can encode those qualities, we can predict what someone wants, needs, and will pay for, at scale - wildly beyond the capabilities of traditional personalization engines. We can also understand the consumer across domains and platforms.
We built PSYKHE AI to formalize that intuition into infrastructure: a psychographic world-model that understands why people prefer what they prefer. A structural layer that lets digital systems reason about human taste with the same depth that good merchants have always operated with instinctually.
Personality - who the person is - is the piece that’s been missing from this space for over two decades. And it’s why we’re building it now."
Everything we’re drawn to - cities, homes, art, cars, music, travel destinations - isn’t just about surface features, but about the deeper qualities an object or place embodies. We seek that which reflects who we are, and who we aspire to be. If we can encode those qualities, we can predict what someone wants, needs, and will pay for, at scale - wildly beyond the capabilities of traditional personalization engines. We can also understand the consumer across domains and platforms.
We built PSYKHE AI to formalize that intuition into infrastructure: a psychographic world-model that understands why people prefer what they prefer. A structural layer that lets digital systems reason about human taste with the same depth that good merchants have always operated with instinctually.
Personality - who the person is - is the piece that’s been missing from this space for over two decades. And it’s why we’re building it now."
Anabel Maldonado
Founder & CEO, PSYKHE AI
