Case study

Salestreak: building a selling workspace with AI product content

Salestreak gives online sellers one place for their products: a catalogue with variants and stock, AI product photos and listing copy in five languages, their own webshops with checkout, publishing to Etsy and print on demand through Printful. It launched in 2026 and is in its early days. This is what's built so far, and how.

Product
Selling workspace for online sellers
Launched
2026
Stack
Rails 8.1, Stripe Connect, OpenAI
Languages
5
Status
Live, early days
One mug photographed on a cluttered counter, the same mug as a clean packshot, and in a styled breakfast scene
One phone photo in, a packshot and a styled scene out. Example images, generated for illustration.
The challenge

The challenge: one product, too many places

Small online sellers do everything themselves. They photograph products on the kitchen counter, write descriptions per channel and per language, keep stock in a spreadsheet and copy orders between their own shop and Etsy. Every new channel multiplies the work.

The aim was one catalogue as the starting point for photos, copy, shops and channels, and one list for every order, with the hard parts (payments, stock and labelling AI content) handled properly from the start.

Our approach

Our approach

Salestreak started on 10 September 2026 from the Twelve Lumens codebase. Billing, the image pipeline, five languages and the deployment came along, already proven in production. On that first day the schema was rebuilt around products, stock, channels and orders, and the first Etsy connection went in.

  • One catalogue as the source for photos, copy, webshops and channels
  • The Twelve Lumens image engine, adapted from rooms to products
  • Payments that are always right: idempotent, row-locked and de-duplicated
  • An MCP connector so AI agents can work in the shop, within the user's own rights

A catalogue built for variants and stock

Products have option axes such as size and colour that generate variants. Every variant carries its own SKU (unique per workspace), EAN, barcode, MPN, price, compare-at price, cost, weight and dimensions in millimetres, plus a stock policy: tracked, untracked or made to order.

Stock is kept per location (warehouse, fulfilment partner, store or supplier), with reservations and a movement ledger that records every change: received, sold, returned, adjusted, damaged or transferred. It's the same groundwork we lay for e-commerce platforms we build for clients.

The catalogue model

  1. Product

    Name, brand, type, tags and a description per language.

  2. Variants

    Generated from options such as size and colour.

    • SKU
    • EAN
    • price
    • weight
  3. Stock per location

    Warehouse, fulfilment partner, store or supplier, with reservations.

  4. Movement ledger

    Every change recorded: received, sold, returned, adjusted, damaged, transferred.

Product photos from a phone snapshot

A vision pass sorts every upload into one of nine shot types and writes captions and alt text in five languages. Declutter turns a snapshot into a clean packshot on a neutral backdrop. Restyle places the product in one of seven scene presets, or in a custom brand scene built from a description and reference images.

Each shot can be presented worn, on a table, as a flat lay, hanging, held or as a packshot, under one of five lighting setups. A hard rule in the prompt makes products rest or hang the way they would in real life.

Every AI render gets a burned-in disclosure bar and IPTC metadata. Listing copy is drafted natively in English, Dutch, German, French and Spanish from the product data and the photo analysis, with an explicit rule never to invent materials.

The same mug in a styled breakfast scene with linen, eucalyptus and a croissant
A speckled ceramic mug on a cluttered kitchen counter
Before After
The same mug before and after a restyle. Example images, generated for illustration.

Webshops with a checkout that's always right

Sellers can run several webshops, each in up to five languages with hreflang. Checkout uses Stripe Checkout, charged directly on the seller's own connected Stripe account, with SEPA and bank transfer, VAT invoices, refunds and dispute handling. Salestreak takes its platform fee through Stripe Connect and refunds it pro rata when the seller refunds. More on that in our Stripe integration work.

Stock is held for an hour during checkout and released by a sweeper job when a cart is abandoned. Checkouts carry idempotency keys, status changes take a row lock and webhook events are de-duplicated, so a retry never charges twice.

The shops also follow strict honesty rules: "Only N left" appears only when one to five units are available, there are no viewer counters, sales counters or countdown timers, prices include VAT and a compare-at price is only shown as a labelled RRP, as the EU Omnibus rules require.

Checkout flow

  1. Cart

    Ship-to country chosen, prices including VAT.

  2. Stock reserved

    Held for one hour and released automatically if the cart is abandoned.

  3. Stripe Checkout

    Charged on the seller's own connected account, with an application fee for the platform.

  4. Webhooks

    Signature checked, de-duplicated and safe to receive twice.

  5. Order and invoice

    Order emails for both sides and a VAT invoice for the buyer.

Etsy and print on demand

Etsy connects over OAuth with PKCE. Publishing creates a draft listing, uploads the images, sets the inventory per variant and then activates it, including Etsy's print-on-demand fields. New orders are fetched every ten minutes and tracking goes back to Etsy.

Printful connects over OAuth too. Designs are uploaded or generated by AI as transparent 2K PNGs, Printful's mockups become product photos, and orders from Etsy or the webshop are routed to Printful, optionally submitted automatically, with webhooks bringing tracking back. We describe the patterns on our page about Etsy and Printful integration.

Other marketplaces and suppliers aren't built yet. Orders from the webshops and Etsy land in one list, with payment and fulfilment status, partial fulfilment and the choice to ship yourself or send the order to the supplier.

An MCP connector for AI agents

Salestreak has its own MCP endpoint, backed by its own OAuth 2.1 server with dynamic client registration and tokens bound to their audience (RFC 8707). Its 37 documented tools cover catalogue, stock, photos, print on demand, websites and orders.

An agent acts with the role of the user who approved it and within the limits of that user's plan, so it can never do more than the person behind it. Read how we approach MCP server development.

How an agent connects

  1. AI assistant

    Claude, ChatGPT or another MCP client.

  2. OAuth 2.1

    Dynamic client registration and tokens bound to the shop audience.

  3. 37 tools

    Catalogue, stock, photos, print on demand, websites and orders.

    • catalogue
    • stock
    • photos
    • orders
  4. The user's role and plan

    Every call runs with the approving user's rights and limits.

Results

Salestreak in numbers

tools for AI agents via MCP
37
languages for shops and AI copy
5
scene presets, plus custom brand scenes
7
automated tests
1,500+

Stack

The same foundation as Twelve Lumens: Rails with Hotwire, Solid Queue in its own container, private image storage on Cloudflare R2 and deploys with Kamal on every push to main.

  • Ruby on Rails 8.1
  • Hotwire
  • ViewComponent
  • Tailwind CSS
  • PostgreSQL
  • Solid Queue
  • Stripe Connect
  • Stripe Tax
  • Etsy API
  • Printful API
  • OpenAI
  • libvips
  • exiftool
  • Cloudflare R2
  • Kamal
What we learned

What we learned

Start from a proven core

Because Salestreak began as a copy of the Twelve Lumens codebase, billing, the image pipeline, five languages and the deployment were tested in production on day one. The first day went into the new domain model, and the credentials were rotated straight away, so the two products share code but no secrets.

Money bugs are the expensive ones

Every checkout carries an idempotency key, status changes take a row lock and webhook events are de-duplicated by their Stripe ID, so retries and redeliveries are harmless. Those paths get tested as carefully as the happy path.

Put the rules in tests

A design-system test scans every view, component and stylesheet: colours from tokens only, the accent colour only in approved places, the wordmark only from its own component, and no em dashes in any language. Rules that live in a test survive deadlines.

Building something for online sellers?

Tell us about your products, channels and payments. You'll get an honest view on what to build, what to buy and where the hard parts are.

martin@balk.dev

Typically a reply within one business day.

Useful to mention

  • What you want to build, in a few sentences
  • Who is going to use it
  • Any deadline or launch date
  • Tools and systems it should connect to