the-tech-art/ai-content-detector

AI Content Detector: Explainable Multi-Vector Forensic Intelligence Suite

Multi-Tier Forensic System
Designed for academic and professional evaluators to detect, quantify, localize, and explain generative AI text across 7 document formats while rigorously preventing false accusations.

---

🌟 Core Architecture & Solution Mapping

Problem Requirement	Evaluator Scenario & Need	VeritasDoc Architectural Solution
1. Ingest Documents	Diverse file formats (.pdf, .docx, .doc, .pptx, .ppt, .rtf, .txt)	Unified parsing pipeline preserving document structural hierarchy and computing SHA-256 checksum.
2. Detect AI Content	Separate machine-generated text from authentic human prose	Tri-Layer Neural Consensus combining Syntactic Discourse (Core Alpha), Generative Cadence (Core Beta), and Adversarial Evasion (Core Gamma).
3. Quantify AI Usage	Mixed/hybrid documents with both human and AI sections	Character-weighted token attribution with 3-sentence Gaussian boundary smoothing ($\tilde{S}i = 0.70 S_i + 0.15 S{i-1} + 0.15 S_{i+1}$).
4. Localize & Explain	Point directly to flagged sentences with transparent reasoning	3-Color Visual Heatmap (Red $\ge 70%$, Yellow $45-69%$, Green $< 45%$) paired with a Diagnostic Explainability Card for every sentence.
5. Minimize False Positives	Avoid wrongful accusations against innocent authors	Platt-scaled calibration, 3-Zone Academic Decision Matrix (Green $<25%$, Yellow $25-65%$, Red $>65%$), ESL Safeguard, and sensitivity modes (Conservative, Balanced, Aggressive).
6. Evaluator UI & Export	Intuitive inspection and tamper-evident audit export	Modern split-screen dashboard with live heatmaps and 1-click SHA-256 hashed PDF audit certificate.

---

🚀 All 4 Bonus Innovators Implemented

1. AI Platform Origin Attribution:
   • Detects generating platform (ChatGPT-4o, Claude 3.5, Google Gemini, Microsoft Copilot) with empirical confidence scores and stylistic token markers (delve, tapestry, furthermore, etc.).
2. Humanizer / QuillBot Evasion Detector:
   • Detects bypass tools (Undetectable.ai, HideMyAI) via Semantic Dissonance Detection and Perturbation Anomaly analysis.
3. LMS Plugin Hook & Chrome Extension (Manifest V3):
   • Canvas / Moodle Webhook endpoint (POST /api/lms/webhook).
   • Ready-to-load Manifest V3 Chrome Extension in extension/ directory for zero-friction browser overlay inspection.
4. Multilingual & Regional Support:
   • Cross-lingual support for Hindi, Spanish, French, and German.

---

🏗️ Project Structure

ai-detector/
├── app/
│   ├── main.py           # FastAPI application, routes (/analyze, /recalculate, /health, /history, /pdf, LMS webhook)
│   ├── config.py         # Zero-configuration defaults from environment (.env)
│   ├── database.py       # Native SQLite (WAL mode) with automatic migrations
│   ├── extractor.py      # Multi-format parser (.pdf, .docx, .doc, .pptx, .ppt, .rtf, .txt) & Hierarchical Tree
│   ├── sanitizer.py      # Zero-width stripping, NFKC normalization, homoglyphs, ligatures & SHA-256
│   ├── detector.py       # Tri-Layer Neural Consensus Engine (Alpha, Beta, Gamma cores)
│   ├── forensics.py      # Perplexity (PPL), Burstiness (B), Zipfian, Stylometrics, ESL Safeguard
│   ├── analyzer.py       # Gaussian smoothing, character weighting, 3-zone matrix, explainability
│   ├── reporter.py       # Tamper-evident ReportLab PDF audit certificates with SHA-256
│   └── utils.py          # Sentence tokenizers & confidence calculation
├── extension/            # Chrome Extension Manifest V3 (Bonus 3)
│   ├── manifest.json     # Manifest V3 configuration
│   ├── content.js        # Floating inspector overlay on any webpage / LMS
│   ├── popup.html        # Evaluator popup inspection UI
│   └── popup.js          # Local consensus inspection client
├── static/
│   ├── index.html        # Evaluator split-screen dashboard with Tailwind CDN
│   └── app.js            # Vanilla JavaScript client (dropzone, heatmaps, diagnostic cards, PDF export)
├── data/                 # Auto-created on startup (persists across reboots)
│   ├── analyses.db       # SQLite database file (WAL mode)
│   └── uploads/          # Saved uploaded documents
├── requirements.txt      # Python dependencies
├── run.bat               # Windows 1-click launcher
├── start.bat             # Alternative Windows launcher
├── main.py               # Root Python launcher
└── README.md

---

⚡ Quickstart Guide

1. Launch the Server (Single Command)

On Windows:

run.bat

or

python main.py

• Web UI: http://localhost:8000
• Health endpoint: http://localhost:8000/health

---

🧪 5-Minute Evaluator Demo Script

1. Step 1: Inspect System Health

   • Open http://localhost:8000/health.
   • Confirm active multi-vector neural cores:

     {
       "status": "ok",
       "model": "Multi-Vector Neural Consensus (3/3 Cores Active)",
       "model_loaded": true,
       "device": "cpu",
       "model_type": "neural_consensus"
     }
     

2. Step 2: 1-Click Instant Evaluation

   • Click 🤖 Pure AI Essay on the dashboard.
   • In 1 second, the radial gauge renders with Red Zone (>65%), highlighting monotonous sentence cadence, low Perplexity ($PPL < 20$), and attributing to ChatGPT-4o (89% confidence).

3. Step 3: Test Interactive Split-Screen & Diagnostic Drawer

   • Click any sentence in the left document text.
   • The Diagnostic Explainability Card on the right instantly inspects that specific sentence, showing its smoothed score, sentence perplexity, clichés found, and linguistic reasoning.

4. Step 4: Export Official Tamper-Evident PDF Certificate

   • Click "Export PDF Certificate" to download the institutional audit certificate containing the cryptographic SHA-256 checksum, 3-Zone Academic Verdict, Neural Consensus Matrix, and Evaluator Disclaimers.