Every collection app has the same problem. The value is in the data, and the data has to be typed in. A serious wine cellar or card collection runs to hundreds of items, and nobody wants to fill in a form hundreds of times.
In two products we built, a photo replaces the form.
Wine labels: Corkprop
Corkprop is a private cellar tool for serious wine collectors. Scan a label and the wine's name, producer, vintage, region and country fill in by themselves.
Recognition is the entry point, not the product. Once a bottle is in the cellar, Corkprop calculates a drinking window for it, so the collector knows what to open now and what to hold. "Surprise me" picks tonight's bottle from the cellar based on the meal and the occasion, and hosts can run tastings where guests score anonymously from their phones.
None of that works if the data is wrong. A vintage read as 2015 instead of 2016 changes the drinking window. That is why a scan should fill the fields for the collector to check, not save a guess silently.
Trading cards: Cortinho
Cortinho is a portfolio manager for trading card collectors. Photograph a card and its name, set, year and number fill in by themselves.
Here the AI does not get the last word. Its answer is checked against the Pokémon TCG database, so the card in the portfolio is a real card with a real set and number, not the model's best guess. That matters because the next step is money: Cortinho pulls minimum, average and maximum sold prices from eBay and shows current value, purchase price and unrealised gain side by side. Collectors can also record PSA, BGS, CGC or SGC grades for each card.
The pattern: recognise, verify, then enrich
Both products follow the same three steps:
- Recognise: a vision model reads the photo and proposes structured fields.
- Verify: the proposal is checked against a trusted source, or shown to the user to confirm.
- Enrich: once the item is known, the app adds what the user came for: drinking windows, live prices, grades, portfolio value.
The verify step is the one that gets skipped in demos. A vision model is very good at reading a label and still occasionally wrong about a digit. Matching against a reference database, or asking the user to glance at the result, turns a good guess into data you can build on.
Design for the hundredth scan
The first scan is a wow moment. The hundredth scan is a chore if the app is slow or makes you fix the same field every time. A few rules help:
- Fill every field you can, and leave unclear ones empty instead of guessing.
- Let the user correct a field in one tap.
- Keep the rest of the interface quiet. Corkprop has no social feed and no gamification. Cortinho uses one accent colour, so gains and losses read at a glance.
Building a product around photos?
Wine and trading cards are two examples. The same approach works for receipts, parts, stock and documents. See our AI integrations and web development work, or tell us what you want to recognise.