Products Hub, Sell-Readiness Scoring & Site Crawler Feeding an AI Sales Bot (Next.js + Supabase)

Arabic/RTL merchant dashboard on Next.js 16 + React 19 + Supabase, with an AI support bot already answering customers over WhatsApp. The store's synced product catalogue was write-only — nothing displayed it, nothing scored it, and the bot could not read it. This project turned the catalogue into a working asset and gave the bot a live source of truth. Delivered: • Products screen — the full catalogue with price, stock, SEO state and readiness, plus server-side search, stock filtering (out / low / in), sorting and pagination. Built for 1000+ products; products whose platform does not report inventory are treated as "unknown", never as zero. • Sell-readiness score — a deterministic 0–100 rating per product across images, description, SEO fields, keyword match and price clarity, stored so the listing can sort and filter by it, and paired with specific Arabic advice ("add a product image", "write a description covering size, material and use"). Runs nightly, plus an on-demand recalculate button so a merchant who has just fixed a product sees the number move immediately. • Website crawler ("connect your site") — the merchant enters a domain and the bot reads the site into its knowledge base, no file uploads. SSRF-guarded fetching with every redirect hop re-validated, robots.txt respected, bounded page and time budgets, content extraction, and incremental sync so unchanged pages are not re-indexed. One crawl per merchant at a time, with visible run history and a sweeper that closes any run that dies. • Bot upgrade — price and availability now come from the live catalogue at the moment the customer asks, never from embedded text that goes stale within the hour. A sold-out product reaches the model with no price attached and an explicit do-not-offer marker, so refusing it is a fact in the data rather than a request in the prompt. Security: every endpoint is session-scoped (an organization id is never accepted from the caller), score writes are revoked from client roles at the database level, and row-level security was verified with a real authenticated token rather than an anonymous key. Acceptance — all verified on the live production site, not a staging copy: - products screen lists the real catalogue; filter and search counts match the database - every product carries a score; recalculating twice does not duplicate or drift - a connected domain appears as bot knowledge and the crawl records completion - the bot quotes the correct live price and refuses to offer an out-of-stock product Tests: 1983 automated tests passing, type-checked and production-built before each deploy.