SkinSense AI

• 

  AI skin analysis interface — upload a selfie to get 8-dimension skin scores and personalized skincare recommendations

• 

  Real-time LLM routing dashboard — tracks fallback events, per-model latency, and cost savings vs GPT-4 (up to 97.7% per request)

What it does

SkinSense AI is a production-grade AI skin analysis platform built on a 3-layer LLM resilience architecture. Users upload a selfie, Qwen Vision analyzes 8 skin dimensions (acne, wrinkles, pores, dark circles, pigmentation, radiance, texture, moisture), and DeepSeek generates personalized skincare recommendations.

The key innovation: when the primary LLM fails, the system automatically switches to Kimi. If both fail, a Rule-Based Engine — built from dermatology guidelines using the actual skin scores — kicks in. Users always get actionable, personalized advice. Zero downtime.

A real-time Dashboard exposes the full routing chain, fallback events, per-model latency, and cost savings vs GPT-4.

How we built it

• Qwen Vision (via OpenRouter) analyzes uploaded photos and returns 8-dimension JSON scores
• DeepSeek generates personalized recommendations as Layer 1
• Kimi serves as Layer 2 automatic fallback (sub-second switch)
• Rule-Based Engine as Layer 3 — always online, zero cost, personalized output based on actual skin data
• Real-time Dashboard tracks every routing decision, latency, and cost
• Built with Next.js 15 + TypeScript, deployed on PM2 + Nginx

Challenges we ran into

• Perfect Corp API free quota ran out mid-development — pivoted to Qwen Vision in hours
• Face++ concurrent request limits on free tier — abandoned and rebuilt the vision layer
• Designing a Rule-Based Engine that gives truly personalized advice (not generic filler) when all LLMs are down
• Making the fallback switching invisible to users while keeping full observability in the Dashboard

Accomplishments that we're proud of

• 3-layer resilience that actually works — tested each failure scenario
• Rule-Based Engine generates personalized recommendations from real skin scores, not static templates
• Real-time cost optimization dashboard showing savings vs GPT-4 (up to 97.7% per request)
• Built and deployed in under 7 hours while simultaneously working on another product

What we learned

• LLM resilience is not just about retries — it requires thoughtful degradation at every layer
• Vision AI and text AI have independent failure modes; both need fallback strategies
• Rule-based systems, when designed well, can deliver genuine value as a last resort
• Cost observability matters as much as uptime in production LLM systems

What's next for SkinSense AI

• Expand to 12 skin dimensions with higher-resolution analysis
• Add product recommendation engine with affiliate integration
• Multi-language support for global markets
• B2B API for beauty brands to embed skin analysis into their own apps

Built With

• deepseek
• kimi
• next.js-15
• pm2
• qwen-vision
• rule-based-engine
• typescript