GlowScore — AI Beauty Analysis App

Overview

GlowScore is a selfie-based beauty and skincare analysis app that we designed and built as a product. A user takes a photo, the app evaluates it against a set of skincare and beauty metrics on the phone itself, and the results come back as scores with a suggested routine. The goal is a tool people return to, so tracking change over weeks matters as much as the first score.

The Challenge

Face-analysis apps face two problems at once. Speed: a spinner while a photo travels to a server and back feels broken in an app about looking your best. Trust: a face image is intimate, and many users hesitate to upload one. There is also tone, because scores can look clinical or judgemental, the wrong register for a beauty audience. We needed analysis that feels instant, keeps the image on the handset for scoring, and sounds like a beauty brand.

Who Uses It and How

The everyday user is building or refining a skincare routine. A typical visit is quick: take a selfie in decent light, read the scores, glance at the routine suggestions and close. Now and then they open the tracking view to see whether a new product seems to help, or compare photos on the before-and-after screen.

A second audience is a beauty brand or partner, who cares about the look and feel and the chance to connect recommendations to a product range.

Our Approach

We chose the model location first, since it drives everything else. Running inference on the phone with TensorFlow Lite means the photo is scored where it was taken, giving near-immediate feedback and limiting how much personal imagery leaves the device. The presentation then centres on encouragement: magazine-style layouts, confident type and animation, and scores framed as a trend to improve rather than a verdict.

What We Built

The capture screen guides the user toward good conditions, since lighting affects any image analysis. The results screen presents scores across several skincare and beauty metrics with short plain-language explanations. A before-and-after screen puts two photos side by side, and a tracking view charts how scores move across sessions. Routine recommendations appear as simple steps.

Architecture and Data

The model runs locally through TensorFlow Lite. Firebase persists result data such as scores, dates and routine choices, so history survives a phone change. Photos for the comparison screen can stay local or be uploaded only with explicit opt-in. Because analysis needs no network, the core loop works on a poor connection, and syncing happens when connectivity returns.

Security, Privacy and Quality

Facial images are sensitive personal data, and in some jurisdictions face-derived data falls under biometric or special-category rules, so consent copy and retention need care. We minimise what is stored, let users delete history and photos, and use Firebase access rules limiting each account to its own records. The app is honest about its limits: cosmetic guidance, not medical assessment. Quality work includes checking model behaviour across skin tones, ages, lighting and cameras, since bias in image models is a known risk, plus testing on low-end and recent phones.

Technology

Flutter gives expressive rendering and animation on Android and iOS from one codebase, which suits a design-led product. TensorFlow Lite is built for compact models on mobile hardware. Firebase supplies accounts, storage for results and a route to add analytics or notifications later.

Outcome

GlowScore gives users quick, private feedback on their skin and a clear record of how it changes over time. The on-device approach removes the upload wait and reduces exposure of personal images, while the editorial design offers beauty brands a distinctive experience.

Frequently asked questions

How much does it cost to build an AI skin analysis app?

The main cost drivers are the model itself (training your own versus adapting an existing one), how many metrics you score, the volume and diversity of image data needed to validate it, and the design and animation effort. Backend needs, such as accounts and history, add further scope. A discovery workshop lets us separate the model risk from the app build.

Is on-device AI better than cloud AI for selfie analysis?

On-device inference gives instant results, works offline and keeps the photo on the phone, which helps with privacy. Cloud models can be larger and easier to update but add latency, running cost and data-handling obligations. Many products combine the two.

Do face analysis apps need special privacy consent?

Often yes. Facial images are personal data, and regulations such as GDPR treat some face-derived data as especially sensitive, so clear consent, purpose limits and deletion options are expected. The exact requirements depend on your users' locations, and legal advice is sensible before launch.