LOCAL PREVIEW View on GitHub

16. Final Summary

The Solution

MangaAssist is an AI-powered shopping assistant embedded directly into Amazon's JP Manga retail store. It combines hybrid AI routing, RAG, LLM generation, recommendation engine integration, and Amazon's existing microservices to create a context-aware chatbot that helps customers discover manga, get answers, track orders, and resolve issues without leaving the store page.

Why It Makes Sense for Amazon

  1. Revenue impact. Customers who get good recommendations are closer to a purchase.
  2. Cost reduction. Deflecting repetitive support questions saves support cost at scale.
  3. Competitive advantage. A domain-specific AI shopping assistant differentiates the JP Manga store.
  4. Data flywheel. Conversations generate intent signals that improve catalog curation and merchandising.
  5. Platform reuse. If MangaAssist works, the same pattern can power other specialty stores.

Why It Improves the JP Manga Experience

Before MangaAssist After MangaAssist
Overwhelmed by 100K+ titles "I like dark fantasy" -> 3 strong picks in seconds
Confused by editions and formats "What's the difference?" -> clear comparison table
Had to search FAQ for return policy "Can I return this?" -> grounded answer
Left the store when stuck Chatbot catches abandoned visitors and re-engages
Waited for human support on simple questions Most common questions resolved instantly
Generic "Customers also bought" Personalized, conversational recommendations

What a Senior Developer Focuses On

1. Architecture That Composes, Not Rebuilds

Amazon already has catalog, recommendation, and order systems. The job is to orchestrate them, not duplicate them.

2. Quality Guardrails From Day One

Wrong prices, hallucinated products, or leaked PII would be launch-ending failures. Guardrails come before scale.

3. Measurable Impact

If the chatbot does not drive revenue or reduce support cost, it should be adjusted or shut down. Metrics and A/B testing are part of the product, not an afterthought.

System at a Glance

graph TB
    subgraph "User"
        U[Customer on JP Manga Store]
    end

    subgraph "Frontend"
        W[Chat Widget<br>React + WebSocket]
    end

    subgraph "Core"
        O[Orchestrator]
        IC[Intent Classifier]
        M[Conversation Memory]
    end

    subgraph "Intelligence"
        RAG[RAG Pipeline]
        REC[Recommendation Engine]
        LLM[LLM<br>Amazon Bedrock]
        GR[Guardrails]
    end

    subgraph "Amazon Services"
        CAT[Product Catalog]
        ORD[Order Service]
        RET[Returns Service]
        PRO[Promotions]
        SUP[Human Support]
    end

    U --> W --> O
    O --> IC --> O
    O --> M
    O --> RAG --> LLM
    O --> REC --> LLM
    O --> CAT
    O --> ORD
    O --> RET
    O --> PRO
    LLM --> GR --> W
    O --> SUP

Key Numbers

Dimension Value
MVP timeline 5 months
Team size (MVP) 10-12
Team size (full) 18-22
P99 latency target < 1.5s for first token
Availability target 99.9%
Conversion lift target +5-8%
Support deflection target 50-70%
Cost per session ~$0.02-0.05
Intents at launch 5 in MVP, 15+ in V3

This system design was prepared from the perspective of a Senior Software Development Engineer at Amazon, focusing on practical, buildable architecture that leverages Amazon's existing infrastructure while delivering a differentiated customer experience for the JP Manga retail store.