The recognition pipeline
Our own model, a suggestion you check
AI Coin Collector runs recognition on its own vision model in the Netherlands and shows the result as a suggestion, not a verdict. You stay in control: check the country and value, and adjust them before saving.
How a scan flows
From the moment you press the shutter to a structured result.
Our own vision model
An open multimodal model on our own AI server in the Netherlands reads the photo and answers in strict JSON: is this a coin at all, country, value, year and mint mark. Location and device data are stripped from the photo first, and no outside AI provider sees it.
You check and adjust
The answer appears as a suggestion. You check the country and value and can correct both before the coin goes into your collection.
Matched to a catalogue
The answer is matched against our coin catalogue on country, denomination, year and mint mark, and looked up on Numista when our catalogue has no match. Recognition needs an internet connection.
Why we built it this way
A suggestion, not a verdict
The app shows the result as a suggestion to check, with no made-up confidence score. When the value is not written on the visible side (the national side of a euro coin), it says the value was estimated from colour and size, and you can correct country and value before saving.
No coin, no answer
If the photo does not show a coin, the app says so instead of inventing one.
Grounded in a catalogue
The value shown is the Numista catalogue price per condition, with a Sheldon grade estimate from your photos. Every value and grade is an estimate, and the app shows it as one.
See the pipeline in action
The grade is an estimate, and the app does not detect striking errors; read how grading works. Join the list and try it on your own coins.