Backlog cleared
0 / 2,481
In June 2026 I shipped a cross-marketplace card matcher for a trading-card arbitrage platform in the US. It runs on one rule: never guess. This page follows a single listing through that rule, decision by decision.
Scroll · one listing, five decisions
Marketplace sellers write titles for humans in a hurry, not for databases. Emoji, bundle offers, shorthand, sometimes the wrong year. Under the noise there is exactly one catalog card being sold.
The platform's margin math starts with knowing which one. Price a listing against the wrong catalog entry and every downstream number lies.
The listing shown is representative, reconstructed for this story. No client data appears on this page.
Normalization first: fold the case, drop the decorations, expand the shorthand the hobby actually uses. NM/M is a condition grade, not part of a name. FOIL is a finish. The giveaway sleeve is noise.
What survives becomes structured fields the catalog can be asked about, instead of a string to fuzzy-match and hope.
The dangerous mismatches are not random cards. They are the same card in a different printing: standard versus alt art, foil versus nonfoil, a 2019 print versus a 2021 reprint. Their prices can differ by an order of magnitude.
So candidate generation pulls every plausible twin on purpose, and makes the scorer tell them apart.
Each signal votes: title tokens, set, finish, printing year, price sanity. This listing agrees with candidate B on almost everything. Almost is the operative word.
The tempting move is to round up. On a 2,481-listing backlog, a matcher that rounds up plants dozens of wrong prices that surface weeks later as bad buys. The bar exists so that temptation has nowhere to go.
This is the rule that won the job: abstain rather than mismatch. Anything under the bar routes to a review queue where a human settles it in seconds, candidates laid out side by side.
The pipeline commits only when it is sure. And it ran read-only against production the whole time; every write stayed with the client's engineer.
The backlog drained: 2,481 listings closed with zero wrong-identity commits. Precision held at 99.4% on the priority stratum and was re-verified after catalog re-scores. Ambiguity became a ten-second human decision instead of a silent error.
“Muhannad is great to work with and is a clear communicator which I value a lot. Will hire him again.”