NYU: CONSUMER PSYCHOLOGY,
BUSINESS/ENTREPRENEURSHIP
+ PHILOSOPHY (AI)
I'M FASCINATED BY THE FRICTION BETWEEN HOW WE THINK AND HOW WE BUILD. AFTER ALL, WE ARE JUST SENTIENT BACTERIA SPACESHIPS TRYING TO NAVIGATE A DIGITAL WORLD. MY WORK IS ABOUT FINDING THE INTUITIVE LOGIC THAT MAKES THESE SYSTEMS FEEL AS NATURAL AS OUR OWN BIOLOGY.
CUSTOMER SUCCESS ENGINEER AT A CONVERSATIONAL AI STARTUP. CO-BUILDING A DISCOVERY LAYER FOR INDEPENDENT FASHION RETAIL. AND GROWING A FEW OTHER THINGS IN THE TREEHOUSE – KEEP SCROLLING. ☺
THE WORK, THE EXPERIMENTS, AND EVERYTHING IN BETWEEN IS FURTHER DOWN. ✦
TWO DOORS AHEAD – MY WORK EXPERIENCE AND MY TREEHOUSE EXPERIMENTS. GO WHEREVER YOU'RE CURIOUS.
Moveo.AI connects three worlds that most enterprises run in isolation – customer service, accounts receivable, and collections – into one shared intelligence layer. In regulated industries like auto finance, insurance, and merchant services, these conversations are auditable, consequential, and full of context that gets lost between silos. I work directly with the founding team at a ~50-person company, spanning both sides of the product: client-facing agent work and internal tooling. I see what clients ask for, how agents actually behave, and where the gap between disconnected systems creates real friction.
I write the behavioral guidelines, guardrails, handover triggers, and escalation flows for production agents – each vertical has different rules for what the agent can say, when it should hand off to a human, and how it handles edge cases. I join client engineering calls, run conversation simulations, and validate each agent before it reaches production.
I built an internal operations dashboard for the customer success team. It gives the team a real-time view of every client account – filtered by geography and region – with a composite health score that flags which accounts are healthy, at risk, or critical based on usage patterns, satisfaction signals, and engagement trends. Before this, there was no single view of where to intervene.
I built an automated red-team framework – one AI runs adversarial conversations that escalate through multiple attack strategies, another scores each response on a severity scale. It tests for prompt injection, jailbreaks, social engineering, and compliance boundary violations. 248+ attack scenarios tested across clients. I packaged it as a Claude skill so any new agent can be stress-tested in minutes instead of days. I also evaluate Moveo’s synthesis LLM for hallucination and response quality – validating model upgrades against structured rubrics before they ship to production.
↗ Customer Success Engineer · Current role.
DRESSX is the world’s largest digital fashion platform, with over 85 million wearables distributed across platforms like Roblox, Snapchat, and Meta, partnerships with brands from Burberry to PUMA, and not a single physical garment produced. I got to work very closely with the co-founder and CEO Daria across two summers in pretty different roles. The first was in marketing and strategy – basically trying to solve the problem of how you sell someone something they can’t touch or wear in any traditional sense. The second was in sustainability research – quantifying the environmental delta between digital and physical garments, and mapping how avatar customization translates into real-world purchasing decisions.
The stat that kept resurfacing across both roles: 84% of Gen Z and Gen Alpha say what they put on their avatars influences what they actually buy.
DRESSX DIGITAL TRY-ON
An 80-page analysis of what people actually buy for their avatars across Roblox, Snapchat, Fortnite, and Meta, along with key insights from DRESSX GenAI. I tracked 30+ brand collaborations and broke the research into two layers: user behavior insights – how people discover, try on, and commit to digital garments – and sales and asset acquisition analysis, mapping which items move, on which platforms, and why. The goal was to surface the patterns that connect avatar purchasing decisions to real-world fashion shifts.
THE REPORT’S THESIS: VIRTUAL FASHION CHOICES ARE PREDICTIVE OF PHYSICAL TRENDS. THE DATA CONFIRMED IT.Each digital garment produces 97% less CO2 than its physical equivalent and saves roughly 3,300 liters of water – enough for one person to drink for 3.5 years. I analyzed 900+ garments to build the environmental case, but the report wasn’t just carbon math. It covered AI’s integration into fashion production and its own environmental footprint, how avatar fashion creates a sustainability pathway that Gen Z actually wants (not one imposed on them), and the behavioral shift where digital self-expression is becoming a social movement that changes what people buy physically.
Ran influencer outreach for the Roblox × Charles & Keith “Summer’s Calling” campaign –10 digital items (shoes + bags), 100+ TikTok and YouTube creators. Grew DRESSX’s presence in the Bitmoji Fashion Marketplace on Snapchat, where 320M+ people dress their avatars daily. Drove adoption of the DRESSX GenAI styling tool – an instant AI try-on that replaced the old 24-hour manual process with text-prompt virtual fitting.
LOS ANGELES, CA · MARKETING & STRATEGY INTERN (SUMMER 2024) → SUSTAINABILITY & REPORTING ANALYST (SUMMER 2025)
Teleported my treehouse from Kyiv to New Jersey for high school, then to New York for NYU. Studying at the intersection of business, technology, and AI — from philosophy to application.
Currently conducting supervised research on why I/O psychology role identity theory predicts LLM behavior when there’s no identity present – why assigning a persona to a system with no self changes its output in ways that parallel how a job title changes a human employee’s behavior, including activating the same stereotypes and moral judgment shifts.
Checkout is a solved problem. Stripe, Klarna, Square, Apple Pay – the infrastructure for moving money after a purchase decision is mature. What isn’t solved is how that decision forms when the product lives in a physical store with no inventory that any system can actually query. Over 80% of U.S. retail is still physical. AI-assisted shopping queries are scaling by orders of magnitude year over year. And the local boutique with exactly what someone is searching for has no way to intercept that intent. An Instagram grid and a Google pin aren’t structured product data. No search engine, no AI agent, no recommendation system can read what’s actually on the rack.
Clara and I spent a year inside Rise, Barclays’ fintech accelerator in the Flatiron District, mapping exactly where that infrastructure breaks down.
Rise gave us structured access to Barclays mentors, their fintech network, and an alumni community of founders who’d already built through similar problems for a year. The first thing we did with it: 300+ consumer surveys and 30+ retailer interviews, face to face in SoHo, Williamsburg, the Lower East Side. Walking into boutiques and asking owners what actually kept them up at night.
Boutiques didn’t need another e-commerce platform. They needed a way to be found by shoppers who were already looking, without having to rebuild their entire digital presence. In our initial outreach, all five boutiques we spoke with wanted the shopper leads, and four expressed willingness to pay – preferring a pay-per-lead model over yet another monthly SaaS subscription. On the consumer side, 50–100+ inspo submissions came in with zero incentives. The recurrent pain point: “I can’t find what I want in NYC stores.”
Platforms like Square and Shopify POS have made BOPIS technically possible for small retailers, but adoption stays low because it still requires real-time inventory accuracy and dedicated fulfillment workflows that most two-to-three-person boutiques can’t maintain. Threadress inverted the model. Reserve without prepayment – hold for four hours, pay in store – doesn’t require the boutique to run an e-commerce operation. It requires them to confirm availability. The payment happens where the boutique already has infrastructure: at the register, in person.
Traditional BOPIS digitizes the entire shopping journey – discovery, selection, payment – and physicalizes only the pickup. Threadress digitalizes only the discovery layer and leaves the full experience – browsing, trying on, purchasing – in the physical store where independent retail actually thrives. The barrier drops from “build an online store” to “confirm you have this item.”
We mapped the competitive landscape across three axes: Discovery Support (Lily AI, Daydream, Stylitics), Channel Expansion (Shopify, SpotOn), and Retail Fintech rails (Stripe, Klarna, Lightspeed). Each axis has players. None of them connect discovery intent to local inventory in a way that works for independent stores – the tools exist in fragments, siloed across platforms, with the merchants who need them most being the least likely to adopt them. Most of the 30+ boutiques we interviewed already had Square or Clover or Shopify on their counter – systems that technically support inventory sync. Not one was using it. The capability existed. The integration didn’t.
Payment rails are universal and interoperable – any card works at any terminal. Discovery infrastructure for local retail is the opposite: platform-specific, high-effort, low-adoption. That asymmetry is what Threadress is trying to fix.
The two big shapes you see are Lorenz and Rössler strange attractors – chaotic math systems that trace butterfly and spiral patterns. I turned them into the heart of the piece. They actually beat: the rhythm comes from your CPU load, and the speed from your battery. If your laptop is working hard, the heart beats faster. If a frame drops, the attractor stutters – a real arrhythmia in the math, not a visual effect.
The curved lines connecting everything are a mycelium network that grows organically between cells. Each connection pulses when the heart beats, so there’s this visible wave of energy that moves through the whole system. Cells are born, live for about two hours, then die and scatter fragments across the cosmos. Meanwhile an LLM is generating text – poems, transmissions from Ukraine, fragments about consciousness – and the machine forgets all of it between generations.
Every two minutes it checks for real air raid alerts in Kyiv. It uses the actual time there, which regions are under alert, how long the sirens have been going. So a transmission generated during a 3am alert reads completely differently from one on a quiet afternoon. The content is never the same twice.
I’m from Kyiv. My dad, my grandparents, my friends – they’re all still there. Since February 2022, air raids have just been part of their daily life. I built this because I kept trying to write about what it feels like to watch an AI generate accurate, well-written text about a war it will never understand – and I couldn’t do it in an essay. It needed to be something you sit with.
The interface is designed like a mission control panel – clocks for Kyiv and New York in the corners, machine vitals at the bottom, thin ruling lines framing the viewport. Only one text layer shows at a time so nothing competes. The cosmos is the thing. Everything around it is just instrumentation.
The majority of retail purchases are digitally influenced. The stores that benefit from that are the ones with structured product data – the large chains with feeds flowing into every search engine and AI system. The hundreds of thousands of independent boutiques carrying inventory that lives on racks, not in databases, are left out. No AI agent, no search engine, no recommendation system can query what’s actually in stock three blocks from you right now.
Clara and I built Threadress to be that infrastructure.
Describe what you’re looking for – or upload a photo. Threadress encodes text and images into the same vector space and matches them against boutique inventory, so “something silky for tonight” and a screenshot from someone’s Instagram both return results ranked by what you actually mean.
Results show what’s in stock, how far it is, and the option to reserve without paying. Hold for four hours, walk in, try it on, pay at the register. The ingestion layer pulls from whatever the store already uses – POS exports, product photos, manual uploads. No structured feeds, no catalog management, no new system to learn.
Free for shoppers. Boutiques pay $55/month for listing, local search visibility, and a demand dashboard – plus $4 per verified reservation. The subscription gives stores something they’ve never had: real data on what people near them are searching for. A boutique that knows 34 people searched for linen pants within three blocks this week makes a different buying decision. The per-lead fee scales with reservation volume and keeps incentives aligned – stores pay for foot traffic that actually showed up.
We’re building geofenced visit confirmation to close the attribution loop beyond reservations, so stores see total Threadress-driven foot traffic. The demand data and the foot traffic data together are the value proposition. Stores that see both keep paying.
Right now, billions of garments sit in independent stores with no digital presence – invisible to every search engine, recommendation system, and AI shopping agent. Threadress is building the structured data layer that makes that inventory queryable for the first time. We started with discovery and reservation. What comes next is a full local commerce graph – real-time availability, demand signals by neighborhood, and the infrastructure for any AI-powered shopping experience to know what’s actually on the rack three blocks away.
We both felt it when we moved to New York – now we’re building something we’d actually use.
ROBLOX · BITMOJI · ZEPETO · AVATAR FASHION ACROSS PLATFORMS
At DRESSX, I explored what fashion looks like when it isn’t physical. I analyzed hundreds of digital garments – following how they lived inside Roblox, Bitmoji, and other avatar spaces, how people styled them, and how much waste they replaced by never being produced.
I contributed to the Digital Fashion Trends Report for Meta, tracking 30+ collaborations and studying how people express identity across Roblox, Snapchat, Zepeto, and Bitmoji.
Slowly, my Treehouse began growing its first digital leaves.
DRESSX ALLOWS ANY DIGITAL
OR PHYSICAL VISION
TO COME TO LIFE
digital closet
avatar ready
Portfolio managers had 24 hours to explain how a Fed rate hike would ripple across 50+ companies. Their tools showed either the numbers or the narrative – never both. Macro shocks move faster than any human can synthesize alone.
NEXUS is an AI co-pilot that fuses LLM narrative reasoning with quantitative modeling, translating macro events into company-level portfolio decisions in real time. It transforms unstructured data – earnings calls, SEC filings, news wires, Fed releases – into predictive signals with the causal chain exposed.
Built with Sagar, Renita, and Vincent for the NYU Stern hackathon, judged by a major Korean fintech company. We won first place.
Fed rate hike → mortgage rates ↑ → housing demand ↓ → homebuilder earnings ↓ → “Recommend underweighting sector.”
The idea behind NEXUS: let the agents handle the research so the GP can focus on the conviction call.
A traditional VC fund with $550M AUM spends roughly $11M a year on management fees –$110M over a fund’s life. Most of that goes to work that is structured, repeatable, and increasingly automatable: pitch deck analysis, financial modeling, market signal scraping, team evaluation, memo drafting. The part that isn’t automatable is taste – sensing cultural momentum, reading founder charisma, having conviction about a market that the data doesn’t fully explain yet.
I built a multi-agent system that replicates the full investment committee process end-to-end. Nine specialized AI agents independently analyze a deal – sentiment and market signals, pitch deck structure, financial metrics, team and founder assessment, and anti-portfolio risk – then higher-order agents simulate partner debate, introduce aesthetic judgment, and synthesize a final decision. The system generates a complete IC memo, a partner debate transcript, and an LP letter in under 60 seconds.
The thesis: most of what a VC does when evaluating a deal (deck analysis, financial modeling, team assessment, market signals) is structured enough for AI to handle. The part that isn’t is the conviction call: backing a founder everyone else passed on, or killing a deal the spreadsheet loves. The system automates everything around that moment and shows you exactly where the agents disagreed and why.
.png)
A browser companion that helps you manage your tabs without thinking about it.
It groups your tabs into Nests, learns what you work on, and pulls up the stuff you’d otherwise lose in 40 open tabs. The pet’s mood reflects how cluttered your browser is, so you get a visual read on your own chaos.
Chrome extension: in review · Website: live
Octopus with Headphones
There are so many tabs. I'm overwhelmed.
Calm things down
Built recently in my Treehouse and now waiting for Chrome's approval.
62 million tons of electronic waste are generated every year. Soil at Agbogbloshie, Ghana’s largest informal recycling site, measures lead at 45 times the EPA safety threshold. Children as young as five strip copper from circuit boards with fire and bare hands. The people who upgrade never see any of this. That distance is the problem.
This project is an AI-generated documentary concept that uses split-screen parallel storytelling to collapse that distance. A single pair of wireless earbuds traced from a New York City upgrader to an e-waste worker thousands of miles away – the purchase and the disassembly, the unboxing and the burning, shown on the same screen at the same time. The form does the argument: you can’t look at one without seeing the other.
There’s a deliberate irony in the medium. The AI that generated these visuals runs on GPUs with a service life of 1–3 years. The hardware rendering this film will eventually end up in a place like the one it depicts. The tool implicates itself – which felt more honest than pretending the camera is neutral.
TINY TREEHOUSE CRYSTAL BALL
Bet On Your Friends (BOYF) — whether you're sitting with your group or miles apart, it's a game for making predictions about each other. Place your guesses, back them with Boyf Credits, and see who actually knows their friends best. My friends and I kept talking about how fun this app would be, so I decided to build it!
TAP TO LOCK IN YOUR CALL ✦
