ProBay: 2,179 products, 964 tests,
and zero orders below cost
ProBay is our own marketplace for digital goods: game currencies, keys, subscriptions, Telegram Stars and Premium. Built on our marketplace engine with a catalogue synced from the pricing system (2,179 products, 130 categories), supplier integrations with HOLD on potential loss, a margin guard verified on 864 orders, a Telegram Mini App, rewards, and an eight-agent AI team (developer, verifier, catalogue, money auditor, support, devops, market analyst, growth).
Why we built it
We had the pricing engine and the catalogue from a client project and a marketplace engine from our own earlier product. Putting them together into a brand we own lets us test every automation on our own money before offering it to clients.
What is in it
Catalogue and pricing. 2,179 products in 130 categories synced idempotently from the pricing system; a separate pricing module with a drift guard against the live store; at launch 1,925 of 1,947 comparable products were cheaper than the main competitor, with a break-even floor and a capped loss only on a small set of deliberately chosen entry products.
Suppliers. Two wholesale integrations with best-price selection, a HOLD mechanism that pauses fulfilment when a supplier price would push an order into loss, and automatic refunds. Catalogue snapshots: 2,887 and 679 positions from the two suppliers.
The guard. A loss sentry checked 864 orders in a verification round: zero below cost. In a later round on live payment channels it removed zero of 2,222 products, meaning the pricing was already safe.
Storefront. Web and a Telegram Mini App with one-click purchase, instant delivery, encrypted stock, rewards (3% cashback, 5% referrals), Telegram Stars with a known markup, legal texts, trust statistics and a reviews system.
The brain. A Claude-based orchestrator with provider fallback to other models, and eight role agents. Nine rounds of development each ended with an adversarial review before merging; the test count grew from 598 to 964.
Status
Live infrastructure, catalogue and payments; preparing the first paid traffic. Internal mechanics (sub-brand network, pricing formulas, supplier terms) are not published.
FAQ
What does the AI agent team actually do?
The developer agent implements tasks and the verifier agent reviews them before a human merges. The catalogue agent keeps products and covers consistent. The money auditor watches every order for margin. The support operator drafts replies. Devops watches the server. The market analyst reads competitor prices. The growth agent proposes experiments. A human approves anything that touches money or customers.