Agentic AI Chatbot - AI Shopping Assistant

AI Shopping Assistant revolutionizes beauty shopping with our agentic AI chatbot powered by cutting-edge Model Context Protocol (MCP) architecture. We solve the industry's $80B return problem.

Comment

Inspiration

AI Shopping Assistant was born from the frustration of online beauty shopping - where 40% return rates plague the industry due to poor recommendations. Traditional systems fail to understand personal preferences like:

  • Skin undertones
  • Texture sensitivities
  • Ingredient allergies The breakthrough came with Model Context Protocol (MCP), enabling AI to deliver human-like consultations with specialized tools for:
    ✔️ Accurate color matching
    ✔️ Personalized product discovery
    ✔️ Context-aware recommendations

What it does

AI Shopping Assistant is an Agentic AI Beauty Shopping Assistant that transforms online beauty shopping by providing personalized, human-like consultations. It solves the industry’s 40% return rate problem by accurately recommending products based on skin tone, texture preferences, and ingredient sensitivities using MCP (Model Context Protocol) architecture for dynamic tool orchestration.

How we built it

Tech Stack

  • Backend: Hono.js, TypeScript, Node.js
  • Databases: MongoDB (products), Weaviate (vectors), Redis (cache)
  • AI/ML: Gemini API, FastText, custom embeddings
  • Infrastructure: Docker, AWS,

Key Phases

  1. Conversation Engine: Hybrid intent detection (FastText + LLM)
  2. MCP Tool Agents: Specialized tools for product search, color matching, and pricing
  3. RAG Integration: Weaviate (semantic search) + MongoDB (structured data)
  4. Optimization: Caching, parallel execution, and progressive enhancement for <200ms responses

Challenges we ran into

We built this project from the ground up. At a time when Python dominated the AI landscape, we took on the challenge of developing a robust MCP-based smart shopping assistant using a JavaScript framework. By successfully delivering this complex and high-performance solution on Bolt, we demonstrated that even demanding AI applications can be efficiently built on this platform—leveraging Bolt's strong foundation to significantly reduce development time.

Highlights:

  1. Performance: Initial responses took 2-3 seconds → optimized to <200ms via hybrid routing and Redis caching
  2. Context Management: Complex user references required session-aware tool orchestration
  3. Data Sync: Merging MongoDB (structured) and Weaviate (semantic) data demanded robust pipelines
  4. Beauty Nuances: Encoding color theory and ingredient compatibility into AI logic
  5. To proof the Bolt.new capability, we did this project in javascript.

Accomplishments that we're proud of

  • Achieved 95%+ recommendation accuracy with sub-50ms fast-path responses
  • Reduced perceived latency by 80% via progressive enhancement
  • Built one of the first production-grade MCP implementations for e-commerce
  • Enabled real-time, consultative beauty shopping at scale

What we learned

  • Hybrid AI > Pure LLMs: FastText + LLMs balanced speed and nuance
  • Conversation ≠ Transaction: Users need guided, adaptive dialogues
  • Tool Orchestration is Hard: MCP required parallel execution + fallbacks
  • Beauty is Hyper-Personal: Success depended on domain-specific knowledge graphs

What's next for Agentic AI Chatbot - BeautyMind

  • Voice/AR Integration: Virtual try-ons via camera-based skin tone analysis
  • Multi-Brand Expansion: Partner with Sephora, Ulta, and indie brands
  • Self-Learning UX: Improve personalization via implicit feedback
  • Enterprise API: Offer BeautyMind as a white-label solution for retailers
  • **Expend on Apparel and other ecommerce domain
Share this project:

Updates

Submission history