What it does: Restock Rescue turns an independent boutique owner's budget and plain-language buying instruction (e.g., "Prioritize earrings, but I need at least one necklace for the display table") into a validated, payable wholesale restock basket in one screen: (1) Ask — the owner enters a budget and priority instruction for Northstar Jewelry Supply (fictional wholesale merchant). (2) Validated basket — Muse Spark 1.3 proposes 3–6 SKUs with pack counts and reasons; the server recomputes every price in integer cents from its own catalog, enforces the budget, the ≥60% earring pack-share and must-include-category constraints, repairs an over-budget basket deterministically (or falls back to a labeled deterministic basket), and issues an HMAC-SHA256 signed quote (15-minute expiry) binding quote ID, lines, and total. (3) Approval + payment — the owner reviews the basket and its "Considered & Skipped" tradeoff log, ticks a human-intent checkbox, and pays the exact signed basket through PayPal Orders v2 sandbox checkout. The capture is re-verified against the signed quote before a green COMPLETED receipt (with real Capture ID) is shown. Every basket is labeled by source: "AI recommendation — server validated", "Deterministic fallback — AI unavailable", or "Sample basket — preset, not AI output". Nothing is presented as AI output that is not.
The problem + target users: Independent fashion-jewelry boutiques restock from wholesalers under time pressure with a fixed budget. They must weigh recent sales velocity against stock on hand across a catalog, protect best-sellers, satisfy display and category needs, and stay in budget — usually in a spreadsheet, from memory, while running the shop. Guessing wrong means stockouts on the best sellers and cash tied up in slow lines. Restock Rescue targets those owners (and small multi-location retailers): it converts a budget and one sentence of intent into an inspectable, priced, payable basket — with the tradeoffs shown, not hidden — in under a minute. All demo data is synthetic; Northstar Jewelry Supply is fictional.
How PayPal and AI are used: PayPal (Orders API v2 + JS SDK): @paypal/react-paypal-js renders the checkout buttons only after the owner ticks the approval checkbox. POST /api/paypal/create-order verifies the HMAC signed quote, recomputes the total from the server catalog, and creates the order with intent: CAPTURE, itemized units (packs × 6), custom_id = quoteId, and idempotency key. POST /api/paypal/capture-order re-verifies the quote, checks the PayPal order's custom_id and amount against it, captures, handles ORDER_ALREADY_CAPTURED by returning the existing receipt, and displays the verified COMPLETED capture. This is a sandbox checkout (not production) — a real PayPal sandbox capture, never real money. AI (Meta Muse Spark 1.3): via the Meta Model API (OpenAI-compatible Chat Completions, response_format JSON-schema structured output; reasoning_effort: "minimal" keeps answers inside the 30s timeout) receives the catalog with sales/stock signals and the parsed constraints, under a 9-rule system prompt that forbids price calculation. The server — never the model — owns money: zod-validated output passes a shared cents-based validator; over-budget output gets one deterministic repair pass (lowest urgency-ratio line reduced/dropped, required category and earring share protected); failure falls back to an explicitly labeled deterministic basket. The pipeline guarantees a valid basket OR a user-visible error state. Trust boundary: the HMAC signed quote proves server issuance and basket integrity only — it does not prove owner approval. The checkbox is the human-intent signal; PayPal buyer approval authorizes the payment.
How it was built: Next.js 14 (App Router), TypeScript, Tailwind CSS, single-screen workbench (Ask → Validated Basket → Skipped → Receipt). Money is integer cents end-to-end (catalog priceCents, budgetCents, totalCents) to eliminate float drift across validation, signing, and PayPal item_total reconciliation. src/lib separates concerns: data.ts (16-SKU synthetic catalog, constraint parser, presets), validate.ts (validator, repair, fallback), quote.ts (HMAC sign/verify, in-memory quote store + rate limiter), ai.ts (Muse Spark recommender via Meta Model API), paypal.ts (OAuth, create/get/capture). Six vitest tests cover Scenario A/B arithmetic ($180.00 / $139.80), deterministic repair, both fallback baskets ($151.80 / $149.40), and quote tampering (signature_mismatch). Deployed on Vercel; no database. Quote store and rate limiter are in-memory only; single-instance demo limitation. Not production.
Challenges faced: Sandbox approval stalls — the PayPal sandbox buyer-approval UI repeatedly hung on a spinner while the order had actually reached APPROVED server-side. Verifying state via GET /v2/checkout/orders/{id} (instead of trusting the UI) let us capture the approved $1.00 gate order and obtain a real COMPLETED Capture ID. Model availability — the original free-tier AI model returned persistent 503s during the build, so the recommender was swapped to Muse Spark 1.3 on the Meta Model API (paid credits) behind the same validator and safety nets, and the deterministic fallback kept the demo valid even with AI down. Money integrity — AI output can't be trusted with prices, so the design rejects any AI-supplied price fields, recomputes totals from the catalog in cents, and signs the result — tampering with a single pack count breaks the signature.
What's next: Persist quotes/orders (the in-memory store is a demo limitation) and add multi-instance-safe rate limiting. Merchant accounts with real catalog/sales imports (CSV / POS integrations) in place of synthetic fixtures. Owner profiles with standing priorities, reorder history, and low-stock alerts. Production PayPal onboarding, payouts/reconciliation reporting, and multi-buyer approval roles.
Third-party SDKs and APIs: PayPal Orders API v2 and PayPal JS SDK (@paypal/react-paypal-js) for sandbox checkout; Meta Model API (Muse Spark 1.3) for basket recommendations; Next.js, React, Tailwind CSS, zod, and vitest as the application/test stack.
Built With
- meta-model-api
- muse-spark
- nextjs
- paypal
- tailwindcss
- typescript
