01The problem
Hong Kong's insurance market has 1,500+ product documents across medical, critical illness, life, savings and annuity categories. Agents struggle to find and compare product details during a client conversation; customers find the jargon overwhelming. No unified platform served both audiences with accurate, compliant answers.
02What I did
Product strategy
- Defined a B2B2C position with three subscription tiers (Free / Pro / Enterprise) and a credit-based referral system
- Compliance-first design: every AI answer cites its source document and page
RAG architecture
- Metadata-tagged knowledge base (company, category, plan code, version date) for precise semantic and keyword retrieval
- Dual-prompt system — professional, strategic tone for agents; plain language for customers
- Smart comparison: pick two or three products and the AI generates structured comparison tables
UX and build
- Conversational interface with role switching between Agent Mode and Customer Mode
- Next.js 14, PostgreSQL + Prisma, NextAuth role-based access, Stripe subscriptions
- Streaming responses via ReadableStream for instant feedback
03The outcome
1,500+ unstructured insurance PDFs became a searchable, cited knowledge base
Product research time reduced from hours to minutes for agents
Full SaaS stack shipped: authentication, subscription billing, referral credits
05Tools & stack
Next.js 14
TypeScript
Tailwind CSS
PostgreSQL
Prisma
Dify.ai (RAG)
NextAuth
Stripe