Josh SolisJS Book an intro callBook a call

Retail / marketplace · Vision + pricing

Collectible ID: Near-duplicate image identification and pricing at catalog scale

A dealer points a phone at a collectible. The app must find the exact printing among thousands of near-identical variants, then attach a real market price, while a customer waits.

How it works

Thousands of near-twins

Each dot is one printing with the same name. Many of them look almost the same.

Each constraint removes candidates: set, then number, then finish, then language.

When one candidate stays, the app attaches a market price. When two stay, the app asks the dealer.

Figure: a cloud of dots, one per printing. Each constraint (name, set, number, finish, language) removes dots until one exact match stays and a price is attached.

In the product

When two matches stay, the dealer decides

  1. The scan matches the name, the set and the number.
  2. Two printings stay. They are different only in the finish.
  3. The app does not guess. It asks the dealer to pick one.
  4. The dealer picks the foil printing. Then the app shows a price.

Simplified screen. The item, the price and the chart are examples.

Animated figure: a phone scans a collectible. Three checks pass: name, set and number. The finish check waits, because two printings stay: foil and non-foil. The dealer taps the foil printing. The finish check passes, and a market price card shows 42.50 dollars, the median of three sources.

36-second video

The identity funnel

  1. Thousands of near-twins.
  2. Constraints remove candidates.
  3. One match gets one price.
Transcript
  1. One name can match thousands of printings.
  2. Each constraint removes candidates.
  3. One exact match remains.
  4. The system shows its confidence.
  5. Three price sources merge into one median.
  6. A wrong match is worse than no match.

The parts

  • Scan, identify, price, then buy, sell, trade or add to inventory, in one loop.
  • The identity and catalog layer: near-duplicate printings resolved by set, number, finish and language, across a 151,260-item catalog.
  • A universal CSV import that maps any seller's spreadsheet to exact catalog items.
Show 5 more
  • Sealed products in the same catalog and pricing flow.
  • One review-queue query went from 90 ms to 3 ms.
  • Multi-source price aggregation and caching.
  • Point-of-sale and cross-listing for show floors.
  • Monorepo: backend, web app and mobile scanner.

By the numbers

151,260
Catalog items
90 ms to 3 ms
Review-queue query
API · Web · Mobile
Surfaces

What I learned

Identity is the hard part. A wrong match is worse than no match, so the system shows its confidence.