Giant StoreGiant Store

Sole engineer: the storefront, the shopping assistant and its retrieval pipeline, checkout and the admin dashboard.

Shopping assistants built on a language model love to recommend products the store does not sell, quote wrong prices and answer in the wrong language. A store also needs the assistant to keep working when the model is slow or unavailable.

A complete storefront built with Next.js 16, React 19 and Tailwind CSS 4 that needs no database or backend setup. Each question to the assistant goes through retrieval: the catalog is split into product chunks, embedded locally with fast hashed embeddings, and only the most relevant products are passed to Gemini with a strict, versioned system prompt that forbids inventing products, matches the shopper's language and returns structured JSON.

The server then sanitizes the answer: every recommended product is validated against the catalog by id, so a hallucinated item can never reach the page. Recommendations arrive as product cards with a one-tap Add. If no key is set or the model is unavailable, a built-in local advisor answers instead.

Next.js App Router with route handlers and a proxy. Cart, wishlist and compare live in persisted Zustand stores; auth and orders in React Context; catalog, accounts and orders are kept in the browser so the store starts with one command. The chat endpoint is hardened with same-origin checks, a body-size limit, JSON-only input, Zod validation and rate limiting. Admin-created products appear instantly in the storefront, the instant search, the recommendation rails and the assistant's index.

Making the assistant useful and trustworthy at the same time: grounding it in the catalog, keeping prices and names exact, answering naturally in Arabic or English, and failing gracefully. A full shopping experience around it had to stay fast and smooth on phones.

Local retrieval with hashed embeddings instead of an external embedding API, so the feature has no extra cost or setup. Server-side validation of every model output instead of trusting the prompt. Motion built with Framer Motion that respects reduced-motion settings.

In real runs on the production build, the assistant planned a $1,800 photography kit around a $1,199 camera, recommended a $199 GPS smartwatch in fluent Arabic, and built a $977 desk setup (a 4K monitor, a desk and a mouse) for a $1,000 budget, all as real product cards. A three-item order of $1,707 with the SAVE20 coupon took $341.40 off, unlocked free shipping and was saved at $1,474.85 with tax, and a product created in the admin went live instantly with its own page. Its Vitest suite has 45 tests across 13 files.