Behavioral scientist and computational neuroscientist who bridges human insight with ML system design. I specialize in proving what's really driving behavior at scale — especially when the obvious answer is wrong.
Most UX researchers tell you what users said. I tell you what's actually causing their behavior — and whether your product hypothesis is right or wrong before you ship. My work lives at the intersection of behavioral science, causal inference, and machine learning, which means I don't just surface insights — I translate them into specifications that engineering teams can build against.
During my PhD at UCLA, I decoded brain activity in real-time and used reinforcement learning to influence human preferences unconsciously. I co-developed TRACE, a machine learning method for extracting behaviorally-relevant neural patterns from fMRI data. That foundation in computational modeling is what allows me to think like both a researcher and an ML engineer — a rare combination in UX research.
At TikTok Shop, I conducted research across a 10–20 million user marketplace that directly redirected multi-million dollar product strategies. My two largest projects both overturned prevailing assumptions: one disproved the team's retention hypothesis and redirected an entire quarterly roadmap; the other fundamentally changed how the recommendation algorithm handled content diversity.
I design research to falsify hypotheses, not confirm them. Every study starts with explicit criteria for what would change stakeholders' minds. This builds trust — because when findings do hold, the evidence is unimpeachable.
High-stakes decisions demand convergent evidence. I combine qualitative depth, survey scale, and behavioral data so that no single method's blind spots can mislead a roadmap. If three sources converge, it's signal — not noise.
Research that stays in a slide deck is worthless. I translate user insights into engineering specifications — defining how ML models should weigh signals, what behavioral metrics should replace proxy KPIs, and where systems break human mental models.
A previous team had attempted to reduce content repetition in TikTok Shop's recommendation feed by implementing blanket diversity filters. The result: GMV dropped. Leadership became risk-averse. The ML team was stuck — they needed to address user complaints about repetitive content, but the only solution they'd tried made the business worse.
I was brought in to understand what was actually happening. The question wasn't "how do we show less repeated content?" — it was "what do users actually mean when they say content feels repetitive?"
I designed a multi-method study to separate the signal. The survey reached 5,000 users with a stratified sample across usage frequency segments. Each participant rated three strategically selected videos: one from a category they'd actively searched, one from a category they'd engaged with but never searched, and one from a category they'd never touched. This let me isolate the effect of relevance on perceived repetition.
I used a mixed-effects linear regression model with repetition rating (1–7 scale) as the dependent variable, relevance level as the primary independent variable, and user ID as a random effect to account for within-user clustering. Fixed effects included user segment, product category, creator frequency, and video view count.
Qualitative interviews and cognitive walkthroughs gave me the "why" beneath the numbers — users defined repetition through attributes (same products, same creators, same aesthetic), not through raw exposure frequency.
Users tolerate seeing the same content when it matches their intent. High-relevance content rated significantly lower on repetition — even when objectively shown more often. Repetition is a signal of broken relevance, not excessive frequency.
This reframed the entire problem. The ML team had been thinking about content diversity as a volume control. My research showed them the real moderator was relevance — and I provided the human-centered definition of "similarity" (based on product attributes, creator identity, and aesthetic category) that their models were missing.
I worked directly with ML engineers to translate the findings into model specifications. Instead of suppressing repeated content uniformly, the algorithm needed to assess whether repeated content matched inferred user intent. I provided the signal hierarchy from my regression: search intent was the strongest predictor of tolerance, followed by engagement history, followed by category affinity. This became the weighting structure for the new relevance-aware diversity controls.
US retention for TikTok Shop was significantly lower than other markets. The team's working hypothesis: users needed more reasons to come back. The planned roadmap was built around frequency-based features — loyalty programs, promotional campaigns, gamification. Multi-million dollar investment was on the table.
I questioned the assumption. Multiple factors could explain low retention, and the frequency hypothesis had never been tested. I reframed: are fundamental barriers preventing users from returning, regardless of how many times they've purchased?
Given the stakes — a multi-million dollar roadmap decision — I chose triangulation. Surveys reached 6,000 users. I conducted 20 depth interviews segmented by purchase frequency. Behavioral log analysis from Data Science provided the third leg.
I established explicit falsification criteria upfront: if barriers declined monotonically with purchase frequency, the frequency hypothesis held. If barriers were flat, it was wrong. This transparency with stakeholders meant they were ready to act on findings in either direction.
For synthesis, I used AI-assisted thematic coding with 100% manual validation and 25% double-coding (inter-rater κ = 0.81). I compared barrier rates across purchase frequency segments and ran a propensity score matching analysis to control for confounding variables.
Barriers were flat across all purchase frequencies. Trust (42%), quality (38%), and logistics (45%) concerns were identical whether a user had made 1 purchase or 15. Frequency doesn't build trust — because every purchase on TikTok Shop is an independent gamble, not part of a trust-building relationship.
The deeper insight came from understanding why. Interviews revealed that users think of TikTok Shop as an impulse and discovery channel — not a shopping destination. They browse for entertainment, stumble on products, and "take a chance." Sellers feel random. Quality expectations are low. Users buy different things each time, so no seller-specific trust accumulates. Each purchase is evaluated independently.
This mental model — "TikTok Shop = impulse gamble, not trusted retailer" — explained everything. You can't drive retention with frequency tactics when the platform's fundamental positioning in users' minds prevents trust from building.
41% of churned users were actively purchasing on competitor platforms. They hadn't stopped shopping — they'd stopped trusting TikTok Shop specifically.