WorkCase study 03 / 03
LEVEL
Private, evidence-aware AI analysis of personal presentation, with realistic directions and a 12-week plan. No attractiveness score.
- Type
- Consumer web app, computer vision
- Model
- Paid preview, then full program
- Users
- Adults who want practical style guidance
- Our role
- Product, design, engineering, launch
- Status
- Live
The problem, and how we approached it.
The problem
Face-analysis apps hand out scores. They rate people, over-claim what a photo can show, and push changes nobody can act on.
We wanted the opposite: specific, reversible advice that is honest about what the evidence supports.
Our approach
LEVEL turns seven guided photos into a visual profile, five realistic directions previewed on the person's own photo, and a 12-week plan that adapts to what they actually do.
Findings must cite the photos they came from. Anything uncertain is withheld, and every preview is validated before a user sees it.
Six pieces, one product.
- Boundaries firstUsers set goals, exclusions, budget and time. Nothing is suggested outside them. 18+ only.
- Guided captureSeven views. Angle, lighting and sharpness are checked on-device before upload.
- Evidence-aware analysisEach observation lists the views it was seen in. Uncertain results are withheld.
- Five directionsEveryday, professional, social, alternative and 90-day, each with an exact change log.
- Adaptive 12-week planDaily missions, weekly photo check-ins and a coach that adjusts to progress.
- Private reportA premium PDF covering methodology, findings, directions and the full protocol.
Under the hood
Release only what the evidence supports.
Capture quality is checked in the browser with on-device computer vision, so a failed photo never uploads. Previews pass validators before release, and are clearly labelled as AI-generated.
- On-device face landmarks with MediaPipe for capture checks
- Preview validators for identity, protected regions, age, skin tone and edit isolation
- Typed data layer with Postgres and Drizzle ORM
- Plan engine with missions, momentum and weekly check-ins
- End-to-end and accessibility tests with Playwright and axe
What it runs on.
Frontend
- TypeScript
- Next.js
- React
- Tailwind CSS
Vision & media
- MediaPipe Vision
- Sharp
Data
- Postgres
- Drizzle ORM
- Supabase
- Zod
Quality
- Playwright
- axe accessibility
- Vitest
- ESLint
Hosting
- Vercel
Inside the product.
Screens from LEVEL's public examples. All people and data shown are synthetic.