~/divyam/stories real-time-waste-classification-edge-ml
the edge · on-device ml
Real-Time Waste Classification
An edge ML waste-classification system trained with MobileNetV2 and deployed as a compact TensorFlow Lite model on Raspberry Pi.
- when
- August 2025
- role
- Edge ML Engineer
- stack
- PythonTensorFlowMobileNetV2TensorFlow LiteRaspberry Pi

Overview
This project classifies waste in real time on edge hardware. I trained a MobileNetV2-based model and optimized it for TensorFlow Lite deployment on Raspberry Pi.
Problem
Waste-classification systems need fast inference and reliable accuracy without depending on heavyweight cloud compute. The model had to be small enough for edge deployment while remaining accurate across multiple waste classes.
Solution
I trained MobileNetV2 on more than 60,000 images with augmentation and class balancing, then converted the final model to TensorFlow Lite for low-latency inference.
The deployment target was practical edge use: a compact model that could run quickly on Raspberry Pi without sending frames to the cloud.
Highlights
- Trained on 60k+ images with augmentation and class balancing
- Reached 97% accuracy across three waste classes
- Converted to a compact TensorFlow Lite model
- Delivered sub-100ms inference on Raspberry Pi
Repository
Outcome
The final model showed that compact edge deployment can support real-time classification while keeping accuracy high enough for a practical waste-sorting workflow.