Real models. Real inference.
Not a black box. Four Hugging Face models, loaded once at startup, fed by a capture-and-simulation pipeline and served from both cloud and edge.
Four models, one verdict
PlantVillage MobileNet
Fast & GeneralLightweight classifier across 38 crop diseases — the default first pass, fast enough to run on-device.
ViT Crop Leaf Diseases
Most AccurateA Vision Transformer reserved for the hard calls — the highest top-1 accuracy in the ensemble.
EfficientNet-B3
Field CropsTuned for staple field crops where throughput and accuracy both matter across large acreage.
YOLO11 Pest Detection
Pest SpecialistReal-time object detection on the Insecta-102 taxonomy — draws boxes around active infestations.
How a field becomes a decision
Drone & Capture
Flight planning and an RGB→NIR translation network — derived from anomaly-detection research — let standard cameras stand in for expensive multispectral rigs.
Simulation Twin
A crop-growth and environment simulation generates training data and a safe testbed for the full pipeline before it ever touches a real field.
Cloud Pipeline
Imagery flows into a scalable analysis pipeline that fuses the modalities, runs the model ensemble and renders results to an online dashboard.
Edge Inference
The same models run on Jetson Orin hardware in the field, so results are available without waiting on connectivity.
Want to see it run?
The interactive demo runs the real ensemble on sample fields — or upload your own image.