INSPIRATION
An AI shopping assistant can understand a product name yet still choose the wrong color, size, or other variant. We built CatalogProof to test whether a product choice is supported by catalog evidence before checkout, and to make any correction reviewable by the seller.
WHAT IT DOES
A seller freezes a shopper question, the intended SKU, and a catalog version. NVIDIA hosted NIM diagnoses the variant evidence; the server checks the cited record, field, quote, and shopper constraints. If a correction is supported, the seller must approve it before it creates a new catalog version. CatalogProof then replays the identical question. A separate buyer-approved PayPal Sandbox checkout is tied to the exact SKU and frozen cart. The app reports paid only after PayPal readback confirms one completed capture matching the saved order details.
The demo uses an original, team-authored fictional tote catalog. Its deliberate Stone/Ecru discrepancy is a software test fixture, not a real merchant issue. The live model selected the intended SKU both before and after correction, so the demo shows evidence review and conflict resolution, not a wrong-to-right SKU switch.
HOW WE BUILT IT
CatalogProof uses Node.js 24, SQLite, and a responsive HTML, CSS, and JavaScript interface. NVIDIA hosted NIM handles server-side product-variant diagnosis. PayPal Orders v2 provides Sandbox checkout. Deterministic server code owns SKU, price, availability, cart totals, and payment state; the model can propose a cited correction but cannot edit the catalog or move money.
CHALLENGES WE RAN INTO
We had to make the app reject plausible but unsupported model answers, including citations that did not exactly match the catalog record or shopper constraints. We also hardened payment verification so approval alone is not treated as payment, mismatched or incomplete capture readbacks fail closed, and retries do not create duplicate captures. Sandbox account and currency setup took troubleshooting before the test purchase succeeded.
ACCOMPLISHMENTS THAT WE'RE PROUD OF
We completed a live NVIDIA diagnosis, seller-approved catalog correction, replay of the unchanged shopper question, and a real PayPal Sandbox purchase for the exact synthetic SKU. The USD 24.00 order and its single capture both read COMPLETED, and CatalogProof confirmed the matching payment. No real money moved. The standard suite passed 66 tests; the live NVIDIA suite passed 68. Our public demo video is 1:54 with English captions.
WHAT WE LEARNED
A fluent model response is not proof: product IDs and facts need exact, checkable evidence, and people should approve catalog changes. Likewise, a buyer’s PayPal approval is not a completed payment; the server must verify the capture. Synthetic tests prove software behavior, not merchant demand, sales lift, or real-world impact.
WHAT'S NEXT FOR CATALOGPROOF
We want to test the query-level workflow with merchant-authorized catalogs, compare its findings against existing feed checks, and learn whether merchants value a separate correctness test. We also want to improve judge access with a durable hosted demo. These are future steps; this demo does not claim validated merchant demand or production readiness.