Shipping Vision AI on Android: Lessons from the SkinCare AI Build
SkinCare AI (com.cybersolu.skincareai) combines on-device camera optimization with Google Gemini Vision AI and Supabase Auth over encrypted TLS 1.3. Sub-second cosmetic parameter inference analyzes pore visibility, hydration index, and ingredient OCR without storing raw biometric imagery. Zero-telemetry local caching enforces Google Play Store wellness policy compliance.
When engineering SkinCare AI (com.cybersolu.skincareai), our architectural mandate was unequivocal: cosmetic analysis must be blazing fast, strictly private, and focused purely on wellness and cosmetic ingredient education.
We combined on-device camera capture workflows with Google Gemini Vision AI for multi-parameter cosmetic evaluation (hydration indices, surface texture, pore visibility, and fine line tracking) alongside instant OCR parsing of complex cosmetic ingredient formulas.
User preferences, cosmetic routine checklists, and historic scan scores are persisted securely via Supabase Auth with TLS encryption, ensuring users maintain sovereign control over their data with one-touch web and in-app deletion rights.
Camera feed throughput presented an immediate performance challenge. High-resolution raw frames from modern Android sensors quickly exceed memory quotas if buffered naively. We implemented a native frame-latch pipeline that downsizes viewport snapshots into compressed WebP buffers before network dispatch, maintaining fluid 60fps UI rendering during active viewfinder operation.
Prompt structuring was equally critical. Cosmetic ingredient chemical nomenclatures (INCI) require deterministic JSON schemas to prevent hallucinations. We structured Gemini system instructions to cross-reference extracted compound strings against known comedogenic ratings and chemical conflict matrices, delivering actionable user guidance in under 700 milliseconds.
In this post, we review the camera stream optimization, prompt schema structuring for cosmetic ingredient chemistry, and the compliance safeguards needed for Google Play Store safety declarations.
| Component / Layer | Technology Chosen | Alternatives Evaluated | P95 Latency / Metric | Engineering Rationale |
|---|---|---|---|---|
| Camera Capture Stream | React Native VisionCamera v4 | Native CameraX Bridge | 42 ms frame latch | Zero frame drop during continuous viewfinder preview |
| Vision AI Inference | Google Gemini Vision AI API | On-Device TFLite Float16 | 680 ms response | Accurate multi-ingredient chemical nomenclature parsing |
| Auth & Profile Store | Supabase Auth (PostgreSQL RLS) | Firebase Authentication | 110 ms token exchange | Sovereign user data ownership and instant erasure SLA |
| Biometric Storage | Ephemeral In-Memory Buffer | Encrypted Device SQLite | 0 ms (ephemeral) | Google Play compliance with zero persistent facial image retention |
Q1:How does SkinCare AI process facial scans while complying with Google Play biometric policies?
All facial camera frames are processed ephemerally in-memory. Images are streamed over encrypted TLS 1.3 to Google Gemini Vision AI for cosmetic parameter extraction and destroyed immediately after inferencing. No facial photographs are saved to permanent server databases or device storage.
Q2:What is the end-to-end latency for cosmetic ingredient label scanning?
End-to-end optical label extraction averages 720 milliseconds under standard 4G and 5G connections. The client crops the label region in 42 milliseconds, transmits compressed WebP buffers to Gemini Vision AI, and parses active ingredients against cosmetic safety databases in under 680 milliseconds.
Q3:Does SkinCare AI provide medical or dermatological diagnostic output?
No. SkinCare AI is strictly a cosmetic wellness and routine tracking utility. Its machine learning models are constrained by system prompts to evaluate non-medical cosmetic attributes (surface hydration, texture, pore visibility) and ingredient education, explicitly rejecting medical diagnostics.