
AI and Computer Vision
Document Detection That Never Leaves the Device
LDED (In-house Development)
A neural network under 5 MB detects document corners entirely in the browser — without image upload, without a backend.
- Corners within 5% of the image diagonal
- 93,0 %
- Overlap with the document area
- 97,4 %
- Model size INT8-quantized
- < 5 MB
- Inference per image via WebGPU
- ~ 60 ms
- PyTorch
- ONNX Runtime Web
- MobileNetV3-Small
- WebGPU
- WASM










