ENVIRONMENTAL INTELLIGENCE • KAPATE CONSULTANCY

Smart Waste Sorter AI

Classify solid waste intelligently across 10 streams. Empower circular economy automation with deep learning and visual explainability.

Start Sorting View Analytics

              96.25%
              Test Accuracy
            

              4.7 ms
              GPU Latency
            

              10
              Waste Streams
            

              Grad-CAM
              Explainable AI
            

EfficientNet-B0 Backbone

Real-Time Computer Vision

AI CLASSIFICATION

Waste Classification Studio

Upload a photograph or capture an item with your webcam for instant neural classification.

Model Checkpoint:

Quick Samples:

Input Item

Recyclable

Plastic

Thermoplastic Recyclables

98.4% Confidence

Recommended Bin: Yellow / Commingled Recycling Bin

Latency 4.70 ms

Hardware CUDA (FP16)

Model EfficientNet-B0 (Two-Stage Transfer)

Preparation & Handling Steps

Rinse residue thoroughly to avoid contaminating paper and cardboard in mixed bins.
Crush containers and bottles to minimize transportation volume.
Check SPI resin identification code (#1 PET and #2 HDPE are widely recycled).

Environmental Impact

Recycling 1 ton of plastic saves ~5,774 kWh of electricity and 16.3 barrels of crude oil.

Biodegradation: 450 - 1,000 years

Probability Distribution (All 10 Classes)

EXPLAINABLE AI

Visualizing Model Decisions

Grad-CAM computes gradients at the final convolutional layer of EfficientNet-B0 to reveal which visual features influenced the decision.

Target Class:

Colormap:

Opacity: 50%

Original Image

Grad-CAM Heatmap

Salience Overlay

Target Hypothesis PLASTIC

Peak Focus Region Center Core

Energy Concentration 78.4%

Attribution Latency 4.70 ms

Warm red and yellow regions indicate spatial evidence that drove the model's confidence toward the target class. Cool blue areas had minimal influence.

ANALYTICS & BENCHMARKS

Model Performance Benchmarks

Empirical test-set evaluation across 3,038 unseen images comparing Baseline and Two-Stage Transfer Learning models.

Test Accuracy 96.25% +0.36% over baseline

Macro F1-Score 95.53% +0.61% over baseline

Inference Latency 4.70 ms NVIDIA GTX 1650 (FP16)

Test Images 3,038 Stratified holdout set

Model Comparison Benchmark

Scroll horizontally to view all metrics

Model Architecture	Test Accuracy	Macro Precision	Macro Recall	Macro F1	Weighted F1	Latency
Baseline EfficientNet-B0	95.89%	94.70%	95.21%	94.93%	95.90%	4.70 ms
Two-Stage Transfer Learning Production	96.25%	95.14%	96.01%	95.53%	96.26%	4.70 ms

Evaluation Charts & Convergence Dynamics

RECYCLING GUIDE

10-Class Waste Protocol

Practical handling instructions, preparation steps, and circular economy facts for each waste category.