Under the hoodA real engine. Not a quiz with a score.
Behind every number there’s a computer-vision model that runs on your iPhone, a reader that understands ingredient lists, and an engine that reasons on evidence and learns from you. Here’s how it works.
01Vision model on the iPhone
A computer-vision model (DINOv2 ViT-S/14) compressed to run on the phone. The face is aligned on 14 landmarks around the eyes, cut into 8 areas, and each area is read 10 times from slightly different crops. Three shots are merged, each weighted by how steady it was.
21 MB on device · 8 areas × 10 views · 3 shots · 0 photos uploaded 02Colour science
Light changes everything, so the scan calibrates on the white of your eyes before reading colour. It works in CIELAB and compares each area with the rest of your face, not with an absolute value: that keeps the same face steady from a bulb to daylight.
CIELAB · white balance on the sclera · area-versus-face differences 03The formula reader
Every ingredient list is parsed line by line against a dictionary of more than 20,000 ingredients. GLOUP estimates where the 1% line falls from the preservatives, then reads each active against the dose it needs, its chemical form and how long it stays on the skin.
20,000+ INCI names · 1% line · dose bands · contact time 04Choosing the actives
61 actives, each with its evidence for each concern: 142 rows linked to published studies. Strong evidence weighs more, lab-only or traditional use weighs less, and where there is no proof the app says so.
61 actives · 142 evidence rows · studies with PMID 05Layering
One wash per routine, thinnest to thickest, vitamin C in the morning, retinoids at night, sunscreen last. Strong actives build up step by step, two weeks per step, and never share an evening.
Order by texture · step-up schedule · recovery nights 06Pairings
Ten pairs of actives checked against the literature: the ones that work better together, the ones that need separate evenings, and the myths, like vitamin C and niacinamide, which comes from a 1963 test tube.
10 verified pairs · within a formula and across your shelf 07A model that reasons
How many evenings of a strong active your skin can take is a Bayesian estimate, updated by your evening answers. Your progress line uses a Kalman filter and only shows a trend when it’s 95% sure of the direction.
Beta-Bernoulli with forgetting · Kalman filter · 95% confidence 08It learns from you
Thirteen needs, from your answers, your scans, the season, your age and your diary. With time the engine learns which of them really move on your skin, starting from the research and shifting towards you only as your data gets solid.
13 needs · hierarchical n-of-1 · never beyond ±30% Every release goes through 82 automated test benches: thousands of simulated people, shelves and scans, checked line by line.
