T1ExisTenZ/ARIS-project

A.R.I.S. — Automated Recycling Identification System

Every dead device still contains value.

This repository contains the complete ML architecture, training scripts, and deployment code for A.R.I.S. — an AI-powered e-waste recovery system.

🔗 Live Demo

Try A.R.I.S. on HuggingFace Spaces →

Architecture

Input Image
    │
    ├─► STAGE A: Device Classifier (SigLIP zero-shot → ConvNeXt-Tiny fine-tuned)
    │   └─► {device_type, confidence}
    │
    ├─► STAGE B: PCB Component Detector (OWLv2 zero-shot → Conditional DETR fine-tuned)
    │   └─► [{label, confidence, bbox}, ...]
    │
    ├─► STAGE C: Hidden Component Predictor (Knowledge Graph)
    │   └─► {hidden_parts: [{name, probability, value}]}
    │
    └─► STAGE D: Repairability Scorer (Hybrid Engine)
        └─► {score: 0-100, label: Reuse/Repair/Recycle}

Files

File	Description
ARIS_ARCHITECTURE.md	Complete architecture document with literature review
train_device_classifier.py	Stage A training script (ConvNeXt-Tiny)
train_pcb_detector.py	Stage B training script (Conditional DETR)
aris_knowledge_graph.py	Stage C hidden component knowledge graph
aris_repairability.py	Stage D repairability score engine
app.py	Full Gradio demo app

Key Models

Stage	Zero-Shot (MVP)	Fine-Tuned (Production)
A: Device Classification	google/siglip-base-patch16-224	facebook/convnext-tiny-224
B: PCB Detection	google/owlv2-base-patch16-ensemble	microsoft/conditional-detr-resnet-50

Key Datasets

Dataset	Task	HF Hub
Francesco/printed-circuit-board	PCB component detection (24 classes)	✅
keremberke/pcb-defect-segmentation	PCB defect detection	✅
Custom (to build)	Device classification (16 classes)	❌

Papers

• PCBDet (2301.09268) — PCB component detection on edge
• DWaste (2510.18513) — Waste classification for mobile
• RT-DETR (2304.08069) — Real-time object detection
• Conditional DETR (2108.06152) — Fast convergence detection
• SigLIP (2303.15343) — Zero-shot image classification
• ZeroWaste (2508.18799) — Semi-supervised waste detection