Ecommerce for Jewellery

Full customer flow — browse, a product page with reviews, a persistent cart, checkout with saved addresses and coupons, Razorpay's hosted widget, order history with retry-on-failure and returns. Behind it, an admin suite: product management with an image pipeline that resizes and re-encodes uploads to WebP, coupons, a returns queue that triggers real refunds, and review moderation.

Both verification paths can legitimately race — the browser can close before /verify runs, and the webhook exists exactly to reconcile that. Whichever request wins the conditional update is the one that counts; the loser is a harmless no-op.

The naive version reads the stock count, checks it in application code, then writes a decrement — a gap where two concurrent buyers can both pass the check for the last unit. This folds the check into the write instead:

await tx.product.updateMany({
  where: { id: item.productId, stock: { gte: item.quantity } },
  data:  { stock: { decrement: item.quantity } },
})
// count === 0 means someone else got there first

Postgres evaluates the WHERE clause atomically at write time, so under a real race only one request matches a row — the other gets count === 0 and a clean out-of-stock error instead of an order that can never ship. The same pattern guards coupon usage limits too.

Honest edge case: stale orders are released back to stock lazily, on the user's next checkout attempt rather than a scheduled sweep — so a payment completing after that window could theoretically arrive late. Rare, and acknowledged in the repo rather than glossed over.

Product and curation pages have a "generate copy" button — an admin types a rough draft or a few keywords, and Claude Haiku returns a title, a Markdown description, and a search-snippet meta description as one structured object, parsed straight into a Zod schema by the Anthropic SDK rather than hoping a JSON blob comes back clean:

response = await client.messages.parse({
  model: 'claude-haiku-4-5',
  system: SYSTEM_PROMPT,
  messages: [{ role: 'user', content: input }],
  output_config: { format: zodOutputFormat(GeneratedCopySchema) },
})

The system prompt is explicit about what it can't do: never invent a material, occasion, or count that isn't in the admin's input, and drop the "Highlights" section entirely rather than pad it when there's nothing concrete to list. A rate limit (15 generations per 10 minutes per admin), a model refusal, and a raw API error each get their own distinct response — none of them silently hand back an empty field.