Lumen runs pretrained vision models on your images and shows their reasoning — attention heatmaps, confidence, and labels. One workbench, many models: skin-lesion analysis, object detection, and more to come.
Every model in the workbench doesn't just answer — it shows its working. Here's what you get on each image.
Classifies dermatoscopic images as benign or malignant using a fine-tuned ResNet-50, with the confidence behind every call.
Point the YOLOv8 detector at any photo and get labeled bounding boxes across 80 everyday object classes in real time.
Grad-CAM overlays reveal exactly which pixels drove the model's decision — so you see where it looked, not just what it said.
Probabilities for each class, shown honestly. A borderline call looks borderline — no false certainty dressed up as a verdict.
Models run on a local FastAPI backend, returning predictions and visual explanations within seconds of upload.
One workbench, many models. New vision tasks slot in as additional tabs — the interface grows as the project does.