I've built product in AI, software, and fashion.
I lead UX engineering at a Fortune 500 company and build AI product on my own time. Before tech, eleven years owning accessories lines at Coach and Calvin Klein. Same job every time: decide what matters, ship.
About Me
I spent eleven years working in fashion. I owned handbag and shoe lines from concept to market: the roadmap, the team, the seasonal calendar, production. Product management, in leather.
Then I taught myself to code. First developer at RPM, a Broadway marketing agency (five people when I started, fifty by the time I left). Built the dev practice from nothing, hired every engineer along the way. Since then, UX engineering at a Fortune 500 company: design systems and AI tooling.
Products
01 Design for machinesUnifying eight design systems into four, rebuilt for AI to consume.
problem
Eight design systems, one per business line. Over time they drifted, from each other and from the brand, and maintaining all eight got expensive.
insight
The next user of a design system is a machine. AI can build an interface on its own. What it can't know is how each business should look and behave - and without that written down, every business gets the same generic screen.
solution
One platform that carries what the machine can't guess. Four systems instead of eight, each business's look and behavior built in. Co-wrote the strategy and made the case to leadership.
outcome
A twelve-person platform team, collapsing eight systems into four. Each business's brand and behavior built in - so the AI tooling generates the right screen, on demand, every time.
8 to 4design systems being unified12people on the platform team1platform AI tooling builds on02 Image GeneratorA self-serve image generator that cut the wait for on-brand images to zero.
problem
A research platform publishing constantly: articles, events, videos, all needing on-brand images, made by people who aren't designers. Every image routed through a designer first. A bottleneck for the creators, busywork for the design team.
insight
Bake the design and brand rules into the tool, and a designer's time isn't necessary. The output is on-brand by construction.
solution
An asset generator stocked with brand-approved imagery, color, and type. Creators type a headline, pick a few images, and download a designed, on-brand graphic. A meme generator, aimed at a brand.
outcome
Creators make their own images the second they need them. Designers get their time back for real design work. The tool holds the brand, automatically.
~5 hrsof design time back every week0minutes waiting in a design queue1visual standard, enforced by the tool03 Taste EngineTeaching AI a specific visual taste from a moodboard, then making it prove it.
problem
AI can generate a hundred designs. It can't tell you which one is right for you. "Make it good" isn't a spec.
insight
The real test for AI is holding a point of view. A moodboard is a point of view you can measure: the colors, the spacing, the patterns that repeat, the rules it never breaks. The pipeline pulls those out and holds the machine to them.
solution
A pipeline that writes the taste down: catalog every reference, measure color, density, and composition, cluster what keeps showing up, and distill a thesis the AI has to follow. Open source on GitHub.
outcome
AI that can take a moodboard as the spec and reliably generate to your taste. This site is the proof: its design direction came out of the engine.
76references distilled into 8 channels and 1 thesisMITopen source, public method1:1this site is built from its output04 An agency that runs itselfA supervisor and domain workers that run my operations without me in the loop.
problem
I had about five hours a week for everything outside my day job. Every system I built to save time needed me to run it, so the time it saved went back into running it.
insight
The failure was always the same shape: one big assistant doing everything, with one trust setting. Give it enough access to be useful and you can't let it act alone. Split it into small workers with separate access and each one can earn autonomy on its own schedule.
solution
A fleet of single-purpose workers, each with its own tools and memory. A supervisor verifies the runs and reports, instead of redoing them. Every agent writes to one shared log. A new worker starts read-only, then earns the right to propose, then to act on its own. I promote them one at a time, so a failure always has one source. Nothing goes out without me.
outcome
The operation runs itself and tells me what happened: seven agents on their own schedules, one daily digest instead of dashboards. One of them built this website. The supervisor caught a two-agent outage that had gone unnoticed for four weeks.
7scheduled agents on their own crons1 digestreplaced the dashboards I used to open4 weekssilent failure the supervisor surfaced on day one