Context
Wilczy Łup sells through its own online shop and through partners who recommend it, such as veterinary clinics and other local businesses. A partner who brings a customer should earn a commission on that customer's orders, and the brand needs to know which partner brought whom.
That takes more than discount codes. Orders have to be attributed to the right partner, repeat purchases have to follow the original referral, refunds have to reverse commissions correctly, and every figure has to be traceable. The shop itself stays where it is; the platform sits beside it and listens.
What we built
- Partner register and onboarding. Partners, their locations and their status, with a manual onboarding checklist.
- QR redirect service. Every QR code points to the platform first, so its destination can change without reprinting. Each scan is logged, with an approximate location and without storing IP addresses.
- Partner sticker sets. Numbered, print-ready sticker sheets, generated in batches. Every sheet is checked automatically: each code must scan, and each number must match its code. A set is assigned to a partner on site, by scanning it.
- Codes and referral links. Partner discount codes and referral links held in a signed cookie that cannot be forged.
- Attribution rules. A code entered at checkout wins over a link. A referral lasts 14 days. A customer stays with the partner for six months from the first paid order. Only new customers count.
- Commission ledger. Append-only: nothing is deleted or overwritten, and every change records who, when and why. Provisional figures during the month, final figures at the monthly close.
- Refund handling. A refund reverses the commission in proportion, never beyond what the order earned.
- Staff panel. A web app that installs on a phone like a native app, with two-factor sign-in, five roles and access enforced in the database itself. Money is hidden from roles that may not see it.
- Shop plugin. A WooCommerce plugin that reports paid orders and refunds to the platform through signed webhooks.
How we built it
The platform runs as three small services on Cloudflare Workers, with one Postgres database. Rules about money are enforced in the database, not only in application code: the ledger is append-only at the database level, and each user reads data under row-level access policies. Every incoming order notification must carry the shop's signature, and no notification is counted twice.
We built it with our AI-assisted method. A lead engineering agent and specialist agents wrote the code; separate read-only reviewer agents checked it, and their findings were fixed before release. Deployment runs through a script that executes the automated tests against production after every deploy.
Where AI is used
- In the product
- None in attribution and commission logic, which follow fixed, tested rules. An AI agent prepares the monthly settlement close for human approval.
- In how it was built
- An AI-agent engineering team with independent read-only reviewers, under human approval.
Engineering facts
- 1,245automated test cases
- 23 June – 2 October 2026build period, from first commit to the current feature set
- ~7 s → ~0.8 sstaff dashboard load time after we rewrote how the database evaluates its access policies, in September 2026
Where it stands
The platform has been running in production since 7 August 2026, with test partners and test orders. The brand's public sales launch has not yet happened, so there are no business results to report. We operate the platform day to day.