E-Waste PCB Component Purity & Cannibalization Index: Project Context, System Architecture & Master Action Plan

1. Project Overview & Business Context

The Component Purity & Cannibalization Index is an edge-AI system built for Kabadiwala Connect (e-waste recycling platform). It detects real-time, on-device PCB cannibalization (the illicit stripping of valuable components like CPU sockets, gold edge connector fingers, IC microchips, and SMD tantalum capacitors) to:

Calculate a dynamic Board Integrity Score (BIS).
Emit 8-language vernacular voice warnings to scrap collectors for fair purchasing price adjustments.
Propagate component integrity data across the Collector → Dealer → Recycler supply chain to catch mid-chain cannibalization fraud.

2. Current Project State & Progress Summary

Category	Current Status	Details & Artifacts
Dataset Size	3,923 Normalized JPEGs	Located in `data/images/` and `data/annotated_auto/images/` (6,864 total tracked). Average resolution: $1315 \times 1029\text{ px}$.
Auto-Annotation	6,377 Draft Bounding Boxes	Auto-labeled across all 4,303 image files via YOLO-World zero-shot model (`pipeline/auto_annotate.py`, `conf=0.015`).
Dataset Spec	YOLO Format Ready	Specification saved to `data/splits/data.yaml`. Summary in `data/annotated_auto/auto_annotation_summary.json`.
Updated Taxonomy	7 Fine-Grained Classes	Defined in `config/taxonomy.yaml` (separates fraud desoldered pads from factory empty pads).
Integrity Engine	Track B Complete & Tested	Purity Engine (`purity_engine/`) with form-factor profiles, scrap grade pricing floors, and 8-language voice alerts.
Pipeline Tests	20 / 20 Passed	Full test suite (`tests/test_purity_assay.py` & `tests/test_pipeline.py`) fully compliant.

3. Architecture & Dual-Track Execution Strategy

Development is decoupled into two parallel tracks linked by a frozen contract (model_contract.md):

       ┌────────────────────────────────────────────────────────┐
       │             model_contract.md (Frozen Schema)          │
       └──────────────────────────┬─────────────────────────────┘
                                  │
         ┌────────────────────────┴────────────────────────┐
         │                                                 │
┌────────▼───────────────────────┐       ┌─────────────────▼─────────────────────┐
│ Track A: Model & Dataset       │       │ Track B: Application & Integration    │
├────────────────────────────────┤       ├───────────────────────────────────────┤
│ • 7-Class Refined Taxonomy     │       │ • `MockModelClient` for Dev/Testing   │
│ • Multi-Scale YOLOv8 Training  │       │ • Integrity Engine (`purity_assay.py`) │
│ • Focal Loss for Desoldered    │       │ • Offline Pre-Recorded Audio Warnings │
│ • TFLite INT8 Quantization     │       │ • Collector Lite & Recycler Portal    │
│ • Real-World Test Evaluation   │       │ • BIS-Delta Fraud Discrepancy UI      │
└────────────────┬───────────────┘       └─────────────────┬─────────────────────┘
                 │                                         │
                 └─────────────────┬───────────────────────┘
                                   ▼
          Swap Mock for Quantized Model (`TFLiteModelClient`)
                         (1-Line Config Change)

4. System Requirements & Architectural Adjustments

Based on technical analysis and field realities, the following system-level changes are enforced:

A. Edge Hardware & Memory Optimization (Low-End Android 2GB–3GB RAM)

Sequential Model Execution: Running the 21-class category model and YOLOv8 micro-model simultaneously triggers browser tab memory crashes.
Fix: Enforce explicit tensor disposal (model.dispose()) after Stage 1 classification before loading Stage 2 micro-component detection.

B. Taxonomy Refinement (Eliminating False Fraud Signals)

Fix: Split generic desoldered footprints into:
1. forcibly_desoldered_pad: Exposed pads with jagged solder peaks, torn traces, or blowtorch/burnt flux residue (fraud signal).
2. factory_unpopulated_pad: Clean silkscreen outline with smooth solder dots left unpopulated from factory (NOT fraud).
3. occluded_region: Footprints covered by heatsinks, metal shields, or thick dust.

C. Dynamic Scrap Grade Pricing Floor

Fix: Replace rigid 35% hardcoded floor with category scrap grade floors: $$\text{Adjusted Price} = \text{Base Price} \times \left( \text{Floor}{\text{category}} + (1 - \text{Floor}{\text{category}}) \times \frac{\text{BIS}}{100} \right)$$
- Server/GPU PCBs: $\text{Floor} = 0.15$
- Desktop PCBs: $\text{Floor} = 0.30$
- Low-grade Power PCBs: $\text{Floor} = 0.60$

D. Multi-Angle 2-Sided Inspection Protocol

Fix: Standardize a 2-photo workflow (Front Top-Down + Reverse Bottom-Down / Edge) in Collector Lite to inspect both sides of the PCB.

E. Supply Chain Fraud Tolerance Bands

Fix: Update Recycler Portal intake discrepancy bands to:
- $< 12%$ BIS Drop $\rightarrow$ Normal capture variance (Pass).
- $12\text{--}25%$ BIS Drop $\rightarrow$ Advisory Warning / Discrepancy Flag.
- $> 25%$ BIS Drop $\rightarrow$ Critical Fraud Alert & CPCB Form 6 PDF Audit Log Entry.

F. Offline Vernacular Audio Reliability

Fix: Standardize pre-recorded .opus audio playback via Service Worker CacheStorage as a fallback when Web Speech API regional voice packs are missing offline.

5. Master Plan of Action (Step-by-Step Roadmap)

Phase 1: Baseline Model Training (Track A)

Auto-annotate dataset with 3,130 draft boxes (pipeline/auto_annotate.py).
Fine-tune baseline YOLOv8n / YOLOv8s on data/annotated_auto/data.yaml (imgsz=960 or 1280).
Evaluate per-class mAP@50 and fraud recall on forcibly_desoldered_pad.
Quantize model to TFLite INT8 ($<15\text{ MB}$, target $<1\text{s}$ latency).

Phase 2: Incremental Harvest & Synthetic Data (Optional Extension)

Execute optional incremental scraper batch (~500–1,000 images) using updated config/queries.yaml.
Apply CutMix / Copy-Paste synthetic desoldered pad data synthesis during training.

Phase 3: Purity Engine Module Build (Track B)

Build purity_engine/model_client.py (MockModelClient & TFLiteModelClient).
Build purity_engine/purity_assay.py (Form-factor baseline matching + dynamic scrap pricing formula).
Build purity_engine/vernacular_voice.py (Web Speech API + .opus fallback).

Phase 4: App Integration & End-to-End Verification

Extend Collector Lite (mobile/kabadiwala-lite.html) with 2-photo workflow & inline BIS display.
Extend Recycler Portal (recycler-dashboard-frontend/) with BIS-delta view ($12% / 25%$ flags).
Run automated test suite (tests/test_purity_assay.py, tests/test_bis_delta_api.py).

6. Key Configuration & Pipeline Commands Quick Reference

Run Automated Tests:
```
.venv/bin/python -m pytest
```

Run Auto-Annotation Script:

.venv/bin/python pipeline/auto_annotate.py

Run Incremental Web Scraper:
```
.venv/bin/python scraper.py harvest
```
Dataset Location: data/annotated_auto/