README.md

# 🔌 E-Waste PCB Component Purity & Cannibalization Index

An Edge-AI system designed for Kabadiwala Connect to detect real-time, on-device PCB component cannibalization (the illicit stripping of valuable components like CPU sockets, gold edge connector fingers, IC microchips, and SMD tantalum capacitors).

The system calculates a dynamic **Board Integrity Score (BIS 0–100)**, applies category scrap grade pricing floors, emits **8-language vernacular voice warnings** to scrap collectors, and tracks intake discrepancy tolerance bands ($<12\%$ pass, $12\text{--}25\%$ advisory, $>25\%$ fraud alert) to stop supply chain fraud.

---

## 📁 Repository & Directory Structure

```
.
├── README.md                                # Master project documentation
├── PROJECT_CONTEXT_AND_ACTION_PLAN.md       # Architectural specification & action plan
├── huggingface_dataset/                     # Formatted Hugging Face Hub dataset folder
│   ├── README.md                            # Dataset Card with YAML metadata
│   ├── data.yaml                            # YOLO dataset configuration
│   ├── train/                               # 5,733 training image & label pairs
│   ├── val/                                 # 1,228 validation image & label pairs
│   └── test/                                # 1,229 test image & label pairs
├── splits/                                  # Local leak-free cluster-isolated splits
├── purity_engine/                           # Integrity Engine & Vernacular Voice Package
│   ├── model_client.py                      # Mock & TFLite INT8 model abstraction
│   ├── board_profiles.json                  # Baseline component expectations & scrap floors
│   ├── purity_assay.py                      # BIS calculation & dynamic pricing formula
│   ├── vernacular_voice.py                  # 8-language voice alert generator (.opus fallback)
│   └── pcb_analyzer.py                      # Unified orchestrator entry point
├── training/                                # Custom YOLOv8 training pipeline
│   ├── train.py                             # Fine-tuning & TFLite export script
│   └── train_config.yaml                    # Copy-paste & class-weighted loss config
├── pipeline/                                # Dataset harvesting, validation & dedup
│   ├── auto_annotate.py                     # Zero-shot YOLO-World auto-annotation engine
│   ├── crawler.py                           # Multi-threaded web harvester (16 threads)
│   ├── validator.py                         # Image integrity & resolution validator
│   ├── processor.py                         # EXIF transpose & image normalizer
│   ├── dedup.py                             # Exact & perceptual hash deduplication
│   └── splitter.py                          # Cluster-aware dataset partitioner
├── prepare_hf_dataset.py                    # Formats splits into huggingface_dataset/
└── tests/                                   # Automated pytest test suite (20 tests)
    └── test_purity_assay.py
```

---

## 📊 Dataset Statistics & Hugging Face Hub CLI Upload

The dataset contains **8,190 clean normalized JPEGs** and **6,377 fine-grained bounding boxes**.

### Dataset Partitioning

| Split | Images & Bounding Box Pairs | Ratio |
| :--- | :--- | :--- |
| **Train** | **5,733 images** | $70\%$ |
| **Validation** | **1,228 images** | $15\%$ |
| **Test** | **1,229 images** | $15\%$ |
| **Total** | **8,190 images** | $100\%$ |

### 🚀 Pushing Dataset to Hugging Face via HF CLI

The dataset is pre-formatted in `huggingface_dataset/` with YAML metadata. Push it directly to your Hugging Face Hub account using the Hugging Face CLI:

```bash
# 1. Login to Hugging Face
huggingface-cli login

# 2. Upload the staged huggingface_dataset directory
cd huggingface_dataset
huggingface-cli upload <your-hf-username>/<your-dataset-name> . --repo-type=dataset
```

---

## 🏷️ 7 Fine-Grained Class Taxonomy

| Class ID | Class Name | Description |
| :--- | :--- | :--- |
| `0` | `cpu_socket` | LGA, PGA, or ZIF processor socket footprints |
| `1` | `gold_finger` | Gold-plated edge connector contacts (PCIe, RAM contacts) |
| `2` | `ic_chip` | Packaged ICs and BGA microchips |
| `3` | `tantalum_capacitor` | SMD tantalum capacitors (molded yellow/black/orange) |
| `4` | `forcibly_desoldered_pad` | Exposed pads with jagged solder, blowtorch marks, torn traces (**fraud signal**) |
| `5` | `factory_unpopulated_pad` | Clean silkscreen outline with smooth solder dots left empty (**NOT fraud**) |
| `6` | `occluded_region` | Footprints obscured by heatsinks, metal shielding, or heavy dust |

---

## 🏋️ Custom Model Training Pipeline

Fine-tune YOLOv8 on the dataset using **Copy-Paste synthetic augmentation** (`copy_paste: 0.8`) and **elevated class loss gain** (`cls: 2.5`):

```bash
# Run training using project virtual environment
.venv/bin/python training/train.py --epochs 50 --imgsz 960 --batch 8
```

### Dynamic Scrap Grade Price Formula

$$\text{Adjusted Price} = \text{Base Price} \times \left( \text{Floor}_{\text{category}} + (1 - \text{Floor}_{\text{category}}) \times \frac{\text{BIS}}{100} \right)$$

* **Server / GPU PCBs**: $\text{Floor} = 15\%$
* **Desktop PCBs**: $\text{Floor} = 30\%$
* **Low-Grade Power PCBs**: $\text{Floor} = 60\%$

---

## 🛡️ Running Integrity Assay Engine (`purity_engine`)

```python
from purity_engine import PCBAnalyzer

analyzer = PCBAnalyzer()

# 2-Photo Workflow (Front + Back PCB Photos)
result = analyzer.analyze_board(
    front_image="front_top_down.jpg",
    back_image="back_bottom_down.jpg",
    base_price=100.0,
    form_factor="desktop_atx",
    lang="hi",  # Hindi voice alert
)

print(f"Board Integrity Score: {result['board_integrity_score']}%")
print(f"Adjusted Purchase Price: ₹{result['adjusted_price']}")
print(f"Voice Alert: {result['voice_alert']['transcript']}")
```

---

## 🧪 Running Automated Unit Tests

```bash
.venv/bin/python -m pytest tests/test_purity_assay.py tests/test_pipeline.py
```
*(All 20 unit tests passing)*