Wastely

Built an AI platform (team of 5) classifying photos of public trash bins as full or empty. Fine-tuned a MobileNetV2 CNN to 85% held-out accuracy (94% precision) and deployed it via ONNX at 3x smaller model size.

• Built an AI web platform that classifies photos of public trash bins as full or empty to help cities prevent illegal dumping. Engineered an 11-feature image pipeline (OpenCV) feeding a SQLite database and a Flask app.
• Benchmarked 3 classification approaches on ~600 labelled images: a transparent 16-rule engine (60% accuracy), a Random Forest (61%), and a fine-tuned MobileNetV2 CNN reaching 85% held-out accuracy and 94% precision using transfer learning and data augmentation.
• Deployed the CNN with ONNX Runtime for CPU-only inference, cutting model size by 67% (58 to 19 MB) in line with the project's Green-IT constraint.
• Shipped as a multilingual Flask web app (46 routes, 4 languages) with geolocated email alerts, QR-code citizen reporting and OSRM-optimised collection routes.