A developer has created an AI-powered shopping assistant for a gift store using the Qwen 3.5 9B model. The system architecture separates logic between a backend built with NestJS, Prisma, and PostgreSQL, and an AI harness that manages model interactions. The frontend is developed using React 19 and Vite.
The assistant operates by querying the store's API to filter products based on age, price, and delivery time. It does not have its own database; instead, it relies on the backend to calculate totals and manage cart state. The model is responsible for selecting up to five items, explaining choices, and understanding user intent, while the code handles budget constraints and discount logic.
A 20% discount is automatically applied if a birthday falls within five working days. The system includes a validation layer where the code verifies all prices and amounts mentioned by the model against the backend data. If the model generates incorrect figures or violates constraints, the system defaults to a template response.
The developer tested the assistant on a catalog of 64 items. Performance metrics show that while the model-only approach is faster, the current architecture—which uses three sequential model calls—is more reliable for maintaining budget accuracy. The source code is available under the MIT license in the nmrcs/cart-agent repository.
