Four food brands, one backend, and books that classify themselves.

Part 1: One backend, isolated by the database

An Express 5 + Prisma API over Postgres uses row-level security to isolate every store’s data in the database itself, not in application code. Owner roles and per-store API keys scope access, and a provisioning script creates a new store with its owner, checkout rules and key in one run. Storefronts are per-brand React/Vite and Next.js apps, and the admin portal is Next.js.

• Tenant isolation enforced by Postgres row-level security
• A new branded store in one provisioning run
• Per-store SEO sitemaps generated at build time

Part 2: Operations that run themselves

ShipStation webhooks update order status automatically, catch voided shipments and record shipping cost per order. TaxJar product tax codes handle tax; a checkout-rules engine, option-level inventory and one-time coupons handle the catalog and promotions. Order confirmation, welcome, password-reset and shipment emails go out through Postmark and SendGrid from each brand’s own sender identity, and a TensorFlow.js service finds the subject of every product photo and crops the variants automatically.

• Fulfillment status without anyone refreshing a dashboard
• Store-branded transactional email for every brand
• Object detection that auto-crops product imagery into Cloud Storage

Part 3: Funnels measured to the purchase

The Knuckle Ball storefront carries a full Meta Pixel e-commerce funnel (page view, product view, add to cart, checkout, and purchase with order value), and the flagship runs Google Tag Manager, so paid traffic can be judged on purchases rather than clicks. Structured data and sitemaps ship with every store.

Part 4: An AI CFO that can look, but cannot touch

Bank and card accounts link through Plaid, and an idempotent cursor engine syncs every transaction into Postgres without double-counting if a page of results fails midway. Claude classifies each merchant into a real chart of accounts: live for new transactions, and through the discounted batch API for up to two years of history pulled on first link. Code, not the model, decides P&L grouping and tax deductibility. A streaming “Ask the CFO” advisor answers from a live snapshot of the books, beside views for burn, payroll, labor cost, payables and shipping cost.

• The AI classifies; deterministic code decides the accounting treatment
• Transfers and card payments de-duplicated so burn is never double-counted
• QuickBooks import and CSV export for the CPA

Part 5: New revenue lines, launched like products

When Craveble opened a fulfillment business for other food brands, we shipped its lead-generation site: SEO-structured pages for frozen, refrigerated and ambient storage, JSON-LD, a sitemap, and inquiries routed to both founders, deployed to Cloud Run with a single command. Its Amazon FBA prep page started as a Bolt.new scaffold, the same AI app builder many founders start with, and our engineers took it from there to Cloud Run.

• A landing page per service line, with structured data
• Lead routing through a health-checked backend
• A deadline-driven page that began in Bolt.new