Case study · Published, later decommissioned

MirrorMirror

A privacy-conscious AI beauty analysis app that combined computer vision, a LLaMA-based retrieval system, and expert knowledge to make specific, explainable recommendations.

Problem & role

Personalization without opacity

The product asked users to take a selfie and returned personalized skincare routines and makeup recommendations. Josh built the system end-to-end under ACON AI: model integration, interface, privacy architecture, and App Store and Google Play deployment.

The challenge was to avoid generic advice. Recommendations needed to identify a product, a shade, and the reason it fit that specific face while handling a highly personal image responsibly.

Technical approach

Vision plus grounded retrieval

Encrypted captureUser photo handled through a privacy-conscious upload flow.

270+ landmarksAnalyze hydration, tone, texture, oiliness, and redness signals.

LLaMA-based RAGRetrieve from influencer tutorials and cosmetologist-validated product knowledge.

Explainable resultReturn product, shade, and rationale; photos auto-delete after 30 days.

Computer visionLLaMARAGiOS & AndroidEncrypted photos

Outcome

A shipped learning system

MirrorMirror was published to both major app stores and was subsequently decommissioned. Its underlying model architecture informed COINAILYZER, carrying forward the pattern of combining deterministic vision work, domain-grounded knowledge, and explanations that help users judge an AI result.

See the COINAILYZER case study Return to projects Connect to the résumé