Skin analysis and advisory

The context

Consumer health, skincare. The client was building a product whose core promise was the assessment itself, not a bolt-on feature: point a phone at your face, get an honest read. The operating environment was the open world: ordinary phone cameras, uncontrolled lighting, every skin tone and age. Constraints were firm: a result in under a minute from one selfie, guidance a lay person can act on, careful handling of anything that looked medical, and no per-user human review at consumer scale.

Why normal software was not enough

No rules engine can read a face. Wrinkles, tone, texture and blemishes are visual, continuous, and different in every photo, so hand-written logic and questionnaire scoring were never options. Only vision models trained and validated on real-world selfies could assess skin from one image, and only a condition-aware advisory layer could turn that assessment into guidance specific to the person, not a generic quiz result.

The problem

Skincare advice was either a dermatologist appointment or a guess at a shop counter. The product goal: point a phone camera at your face and get an honest, personalized read in the time it takes to take a selfie.

What we built

We built vision models that assessed skin condition from a single selfie, wrinkles, tone, texture, blemishes, distilled into a relatable skin age, then paired the assessment with a personalized advisory engine mapping conditions to routines and products. Edge cases and sensitive results were phrased with care and routed toward professional advice.

What production made hard

• Selfies arrive ugly: bad lighting, odd angles, makeup, filters and low-end cameras. The models had to hold up on those images, not on studio captures.
• Consistency across skin tones was a requirement, not an average. Training and validation data had to be deliberately balanced so accuracy did not quietly concentrate on some complexions.
• The line between cosmetic and medical is thin. Sensitive results had to be phrased with care and routed toward professional advice, a product-policy decision as much as a modeling one.
• Repeatability: users retest. Two selfies of the same face in similar conditions had to land on a similar skin age, or trust in the whole product collapses.
• Everything, capture to assessment to personalized advice, had to fit inside the under-a-minute promise on ordinary phones.

What a similar project needs

• Real-world image data spanning skin tones, ages, lighting and cameras, with condition labels checked against domain experts. A clean studio dataset alone will not survive contact with users.
• A timeline shaped like ours: a short prototype sprint proving the core assessment on real images, then a longer hardening phase dominated by in-the-wild image quality, fairness validation and advisory tuning.
• A small mixed team: vision engineers, a product or mobile engineer, and domain expertise to define labels, thresholds and the advisory rules.
• A decision, made early, on what counts as sensitive: which results get softened phrasing and a nudge toward professionals. That boundary needs sign-off from whoever owns legal and brand risk.
• Eyes on the real risks: overclaiming into medical territory, scores that wobble between retakes, and skin-tone bias found after launch instead of before. All three are testable before shipping.