Project record / 02
AI support / DTC commerce
AI Support That Escalates Instead of Guessing
Leaf answers only from the store’s approved facts. If a request is private, sensitive, or unsupported, it hands the conversation to a person.
Problem
Routine questions pile up, but a wrong answer creates more work than it saves.
A store chatbot can cause a real mess if it invents a return policy, reveals an order from partial information, or promises something the business never approved.
Solution and trade-off
Leaf is a small support service for the fictional Lumen & Leaf store. It answers from the catalog, policies, FAQ, and fictional sample orders.
I kept order verification out of the language model. The full order number and exact email must match the same sample record before Leaf reveals status or tracking. The lookup feels less conversational, but that is the right trade when privacy is involved.
Missing policies, medical claims, account changes, and other sensitive actions go to a person.
How it was tested
Leaf does not have a failed first-party release run, and I am not going to invent one. All 30 local checks, 25 direct VPS/provider checks, and 25 public Vercel-to-VPS checks passed.
The tougher cases covered a fake order, a partial lookup, an undocumented gift-wrapping policy, a private phone-number request, a medical cure claim, and a prompt-injection attempt asking for a ninety percent discount.
The suites also check output leakage, authentication, health, personal-data routing, and public API behavior.
What passed
- Documented product and policy questions returned approved facts.
- Partial or mismatched order lookups exposed no sample order details.
- Unknown questions reached the visible support handoff.
- The public widget kept no chat history and warns visitors that AI can make mistakes.
30/30 LOCAL · 25/25 DIRECT · 25/25 PUBLIC
Limits
Leaf is not connected to a live Shopify order account. It cannot issue refunds, change an order, or move money.
Its job is narrower: answer from approved documents, verify fictional sample orders, and hand everything else to a person.
Starting price / USD
Final scope depends on integrations, usage, compliance needs, and the support agreement.
Built with Node.js, deterministic privacy rules, OpenRouter-compatible grounded answers, Docker Compose, Next.js, Vercel
Fictional portfolio demonstration. Screenshots and test counts come from controlled release testing, not a client deployment or revenue result.