Result: 10% increase in site conversion rate
A/B tested with triangulation validation. Built in 2 months.
Overview
Koffee Kult is a large-scale coffee roaster running on Shopify. They needed a way to reduce customer support ticket volume while improving the shopping experience. I built KKChat — a RAG-powered AI chat agent that lives on their storefront, handles product questions, order status inquiries, and subscription management in real time.
The patterns here — production RAG, deterministic sub-flows for stateful operations, eval-driven iteration — apply directly to B2B SaaS contexts where engineering teams need to ship AI features against real catalogs, real users, and real cost ceilings.
The Challenge
Koffee Kult’s support team was fielding repetitive questions about products, shipping, and subscriptions. These were high-volume, low-complexity interactions that didn’t need a human — but the answers required real knowledge of their product catalog, shipping policies, and subscription system.
What I Built
- RAG-powered chat agent backed by Claude, with vector search across Koffee Kult’s product guides and company documentation
- Deterministic sub-flow for subscription management — structured operations like pausing, resuming, or modifying subscriptions use deterministic logic, not LLM inference
- Shopify MCP integration for product browsing and add-to-cart directly from the chat interface
- In-chat product UI that surfaces relevant products inline based on the user’s message, with direct add-to-cart support — no need to leave the conversation to shop
- Document re-ranking for more relevant responses across their extensive product catalog
- Human escalation flows that seamlessly hand off complex issues to the support team
- Custom iframe chat client designed for security and portability across their storefront
Results
- 10% increase in site conversion rate — measured via A/B testing with triangulation validation
- Reduced customer support ticket volume for common product and order questions
- Ongoing optimization through A/B testing and response quality monitoring
Technical Stack
Built on n8n for workflow orchestration, Claude API for language understanding, vector databases for RAG retrieval, and Shopify’s API for product and cart operations. The chat client is a custom iframe implementation for maximum security and portability.
Have a similar problem?
If your storefront, product, or internal tool has a similar shape — high-volume questions, RAG over real data, deterministic operations that shouldn’t be LLM calls — let’s talk about how the same patterns would apply.