Three Core Technical Pillars
Multispectral Sensing
RGB + NIR cameras capture surface reflectance, polymer transparency, and packaging geometry under calibrated strobe lighting.
Edge Neural Inference
Optimized CNNs perform real-time bounding-box detection and multi-class polymer labeling in under 20ms on edge GPUs.
Spatial Tracking
Rotary encoders synchronize belt speed with camera triggers, tracking each item’s exact trajectory to the ejection nozzle.
Why Local Datasets Matter
Generic Western-trained models fail in Nepal. Local FMCG packaging features distinct domestic brand typography, sachet pouch formats, bilingual Devanagari scripts, and severe deformation from manual crushing.
Local Brands: Domestic beverage, cooking oil, dairy, and snack packaging common in Nepali households.
Deformation Invariance: Models trained specifically on crushed, flattened, partially torn, and folded containers.
Active Learning: Edge devices log low-confidence items for continuous weekly retraining cycles.
All technical specifications are proposed targets — validation is ongoing through our prototype development roadmap.