A multi-stage, institutional-grade platform for detecting plagiarism and AI-generated content in academic project submissions — combining n-gram fingerprinting, semantic embeddings, heuristic analysis, and a multi-class ML classifier into a single auditable pipeline.

1. Project Overview
2. Key Features
3. System Architecture
4. Technology Stack
5. Detection Pipeline
6. AI Detection Engine
7. Database Schema
8. Application Modules
9. API Reference
10. User Roles & Access Control
11. Reporting & Compliance
12. Installation & Setup
13. Configuration
14. High-Performance Mode (Bulk Processing)
15. Admin Service Management
16. Multi-Language Support
17. Subscription & Credits System
18. Known Limitations & Ongoing Work
19. Project Structure
20. License

The WACP Advanced Hybrid Plagiarism & AI-Detection System is a web-based academic integrity platform built for institutions such as universities, colleges, and research organizations. It is designed to analyze student project submissions (dissertations, proposals, reports) and produce auditable, explainable similarity and AI-authorship reports.

Traditional plagiarism checkers suffer from three critical problems:

This system eliminates all three through a hybrid multi-stage pipeline with proper reference flagging, project lineage tracking, and a 4-class AI classifier.

• N-gram Fingerprinting — exact sentence-level matching via Rabin/Karp rolling hash fingerprints
• Semantic Similarity — sentence-transformer embeddings for paraphrase and near-duplicate detection
• Sliding Window Segmentation — 150–300 word segments analysed independently for section-level precision
• Reference Isolation — submissions are compared only against verified reference documents (is_reference = TRUE)
• Project Lineage Tracking — project_group_id ensures re-submissions never inflate scores against earlier versions
• Preliminary Page Exclusion — cover page, abstract, table of contents, bibliography, and references are stripped before analysis
• Multi-format Support — PDF, DOCX, and TXT files are all supported with text extraction and positional mapping

• Heuristic Feature Extraction — 50–200 linguistic, structural, lexical, and stylometric features per document
• 4-Class ML Classifier — distinguishes Human-written, Raw AI-generated, Human-edited AI, and AI-paraphrased text
• Probabilistic Output — results expressed as likelihood ("Likely AI-paraphrased, Confidence: 71%"), never as absolute judgements
• QuillBot / paraphrasing detection — rare-synonym overuse and sentence uniformity heuristics target common bypass attempts

• PDF Highlighting — matches overlaid directly on the original PDF using PDF.js with bounding-box coordinates
• Section-level Reports — each 150–300 word segment has its own similarity and AI-likelihood score
• Dual-audience Emails — admins receive the full analytical report; students receive a highlighted summary without internal model details
• Institutional Audit Trail — all reports are stored for compliance review

• Celery Task Queue — reliable background processing with auto-retry
• Redis Caching — fingerprints and embeddings are cached to skip redundant computation
• Parallel Processing — ThreadPoolExecutor with up to 8 workers
• Bulk Upload — 100+ files processed in 5–10 minutes (vs. ~50 minutes sequential)

┌───────────────────────────────────────────────────────────┐
│                     User's Browser                        │
└───────────────────┬───────────────────────────────────────┘
                    │ HTTP
┌───────────────────▼───────────────────────────────────────┐
│           CodeIgniter 4 (PHP 8.1+) — Frontend             │
│  Auth · Projects · Reports · Admin · Subscriptions        │
└───────────────────┬───────────────────────────────────────┘
                    │ REST (JSON)
┌───────────────────▼───────────────────────────────────────┐
│           FastAPI (Python) — Analysis Backend             │
│  v1 API · v2 High-Performance API                         │
│  Plagiarism Detector · AI Detector · PDF Extractor        │
└──────┬──────────────────────┬────────────────────────┬────┘
       │                      │                        │
┌──────▼──────┐   ┌───────────▼──────────┐  ┌─────────▼────┐
│   MySQL     │   │  Redis (Cache/Queue) │  │    Celery    │
│  (Primary   │   │  fingerprints,       │  │  Workers     │
│   Storage)  │   │  embeddings, jobs    │  │  (3 queues)  │
└─────────────┘   └──────────────────────┘  └──────────────┘

Every submission goes through the following sequential stages:

Stage 1 — Upload & Submission
    └─ Project saved with is_reference = FALSE
    └─ project_group_id assigned (new or linked to existing group)

Stage 2 — Text Extraction & Cleaning
    └─ PDF/DOCX/TXT text extracted with positional data (page, bounding box)
    └─ Preliminary pages removed (cover, abstract, ToC, acknowledgements)
    └─ Back matter removed (references, bibliography, appendices)
    └─ Clean body text passed to analysis

Stage 3 — Plagiarism Analysis
    └─ N-gram fingerprints generated and cached in Redis
    └─ Sentence embeddings generated and cached in Redis
    └─ Compared ONLY against: is_reference = TRUE AND project_group_id ≠ current
    └─ Section-level (150–300 word window) similarity scores computed
    └─ Bounding-box coordinates stored for PDF highlighting

Stage 4 — AI Detection
    └─ 50–200 heuristic features extracted (linguistic, structural, lexical, stylometric)
    └─ Feature vector passed to 4-class ML classifier
    └─ Hybrid Decision Layer combines ML probabilities + heuristic confidence
    └─ Output: class label + confidence score (expressed probabilistically)

Stage 5 — Report Generation
    └─ Auditable JSON report created with all segment scores
    └─ PDF highlight data stored (page, bbox, matched text, score)
    └─ Report saved for institutional audit

Stage 6 — Email Notification
    └─ Admin: full report with confidence metrics and model details
    └─ Student: highlighted submission + summary (no internal model details)

Stage 7 — Post-Analysis Update
    └─ UPDATE projects SET is_reference = TRUE WHERE id = :project_id
    └─ Submission becomes a reference for future comparisons

-- During analysis, only compare against:
SELECT * FROM projects
WHERE is_reference = TRUE
  AND project_group_id != :current_project_group;

This means:

• is_reference prevents an active submission from matching itself
• project_group_id prevents v1, v2, v3 of the same project from flagging each other
• After analysis, the submission becomes a reference for the next student's submission

Recommended models: Logistic Regression (baseline) → Random Forest → Gradient Boosting (production). Deep neural networks are intentionally avoided for legal defensibility and explainability.

The final result combines:

• ML class probabilities (from the trained classifier)
• Heuristic confidence scores (from the feature extractor)

Output example: "Classification: Likely AI-paraphrased, Confidence: 71%"

The following fields are essential to accurate scoring (their absence causes inflated or unreliable results):