I've spent the better part of six years inside the digital product machinery of indie and prestige beauty brands in New York. That meant evaluating software, testing integrations, and fielding the question that came up every single year without fail: when are we adding virtual try-on?
The honest answer, most of the time, was: not yet. And the reason was almost always one of three things — the platform was priced for an enterprise rollout that a forty-person brand could never justify, the implementation required a three-month onboarding, or the shade accuracy fell apart the moment a customer with a deeper skin tone tried to use it. Usually all three.
Lumeglint started as my personal attempt to solve that exact problem.
The gap I kept running into
When I was leading digital product at a prestige color cosmetics house in Manhattan — overseeing the DTC storefront and running evaluations of every AR vendor that came through the door — the pattern became impossible to ignore. The tools that actually worked at production scale were built for accounts with six-figure annual contracts and dedicated technical teams to maintain the integration. The tools that were affordable were either browser toys or so undertone-inaccurate they were functionally useless for complexion products.
What struck me most was the specificity of the failure mode. It wasn't that AR try-on didn't work — it clearly did, in controlled settings, for lighter shades. The problem was that the rendering pipelines had been trained and calibrated predominantly on lighter Fitzpatrick scale values. Fitzpatrick Types V and VI — medium-brown to deep brown skin — were effectively afterthoughts in most of the systems I evaluated. Foundation undertone classification was particularly bad: a cool-toned shade on a warm undertone would render as a near-match rather than flagging the mismatch.
For a brand that had built its identity around shade inclusivity, that wasn't a minor bug. It was a brand liability.
The decision to build rather than consult
After I left that role in late 2023, I spent a few months doing independent digital strategy consulting for indie cosmetics brands — the kind of DTC businesses doing two to fifteen million in annual revenue, where the founder is also the head of marketing and the digital team is one person. These were exactly the brands that needed try-on capability the most, and exactly the brands that were locked out of the existing solutions.
I started scoping what it would actually take to build the infrastructure layer from scratch, prioritizing accuracy across the full Fitzpatrick range from the beginning rather than as a retrofit. I connected with Marcus Webb, who had been doing real-time face segmentation work at a computer vision research lab — his background was in the rendering pipeline, not beauty specifically, which turned out to be exactly right. Building for accuracy-first required thinking about the problem as a color science and computational vision challenge, not a makeup simulation one.
We wrote the first version of what would become LumeCore over the winter of 2024, with a deliberate constraint: every rendering decision had to be validated against skin tone diversity before it shipped, not after.
What the first version taught us
The first working prototype exposed something we hadn't fully anticipated: lighting normalization was a harder problem than shade rendering. A brand's product photography is shot under controlled studio conditions. A shopper using AR try-on is standing in whatever light their apartment or office happens to have. Compensating for ambient light variance — warm incandescent, cool fluorescent, window daylight — without degrading the shade appearance required a calibration layer that we had initially underestimated.
We're not saying lighting normalization is an unsolved problem in computer vision — there's significant prior art. What we found is that applying it correctly in the context of cosmetics rendering required calibrating specifically against pigment behavior, not just general image correction. A translucent lipstick and an opaque matte finish behave very differently under the same lighting shift. That detail took a full month of tuning to get right.
Soo-Yeon Park joined us in mid-2024 to lead brand partnerships. Her background — working directly with indie and mid-market brands on digital strategy — gave us a ground-level view of what the onboarding and catalog-connection workflow actually needed to look like to be usable without a dedicated technical team. The answer, we quickly learned, was: considerably simpler than any existing enterprise solution offered.
Why now, and why here
The timing matters. Beauty e-commerce has structurally shifted toward DTC channels over the past several years, and the return rate problem — driven largely by shade mismatch — has become a meaningful P&L issue rather than just a customer service headache. Brands that used to write off a 20–25% return rate as a cost of doing business online are now being pushed to address it as a sustainability issue, a margin issue, and a customer lifetime value issue all at once.
We closed our first angel round in January 2026, which gave us the runway to complete the LumeCore platform and reach the first cohort of paying brands. The round was small enough to keep us disciplined — three people, a clear technical roadmap, and a pricing structure that works for indie brands from day one, not just enterprise accounts.
The Garment District address is deliberate. We're in the middle of New York's fashion and beauty industry infrastructure, a block from the showrooms and buying offices that still anchor the physical side of the beauty trade. The proximity to that world keeps us honest about what brands actually need versus what sounds good in a pitch deck.
What Lumeglint is and isn't
We're not a consumer app. We're not trying to be a Snapchat filter or a social commerce tool. Lumeglint is infrastructure — a drop-in SDK that beauty brands embed into their own storefronts so their shoppers can try shades before they buy. The brand owns the customer relationship, the data, and the experience. We provide the rendering engine.
The brands we're building for are the ones with real shade ranges — twenty to one hundred fifty SKUs, spanning light to deep complexions — who need a try-on layer that is accurate enough to actually reduce returns rather than just add a feature checkmark to the product page. That's a higher bar than most of the market has held itself to. It's the bar we think matters.
If that sounds like the problem you've been trying to solve for your brand, I'd genuinely like to talk. My contact is on the site.