Our first model Demir-1 is live
We build AI in Türkiye.
We're an independent AI lab. We design, train and ship models here.
Demir-1 in numbers
Our first model recognizes 63 rocks and minerals from a single photo. These numbers come from the test set; training data never leaked into it.
81.15%
Test accuracy
63
Classes
37,800+
Training images
125 MB
Model size
How we measure
We measure every model the same way: training and test images are split by MD5, then we keep the test rocks in hand and shoot them at different angles and light. We publish the result as it is.
Test accuracy is 81.15%, validation 83.12%. Every per-class result is open as CSV.
See test resultsWhat's next
- Grow the class count beyond 63
- Add English to the outputs
- Add field notes, location and hardness
- Move into areas beyond rocks