Kaggle × Google AI Agents: Intensive Vibe Coding Capstone Project
AI-Powered Multi-Agent Misinformation Detection for Modern Digital Platforms
TruthLens is an enterprise-grade, multi-agent fact-checking system designed to help users verify claims spreading across social media (WhatsApp, Telegram, Twitter, LinkedIn). It orchestrates 7 specialized AI agents through a deterministic pipeline to analyze content credibility, bias, and consensus.
Built with: Google Agent Development Kit (ADK) 2.0 · Gemini 2.5 Flash · FastMCP · Gradio · SQLite
| Feature | Description |
|---|---|
| 🤖 7 Agent Pipeline | ClaimExtractor → EvidenceHunter → FactChecker → CredibilityAnalyzer → BiasAnalyzer → VerdictAgent → ReportGenerator |
| 🔗 MCP Protocol | Standalone FastMCP server exposing tools to any compatible host |
| 🖼️ Multimodal OCR | Verify WhatsApp/Telegram screenshots via Gemini Vision |
| 🛡️ Security Stack | SSRF protection, XSS escaping, prompt injection detection, rate limiting |
| 💾 Smart Caching | SQLite with SHA-256 hashing, 7-day TTL, WAL mode for thread safety |
| ✅ Evaluation Suite | 24 automated tests across 5 categories — 100% pass rate |
| 📊 Structured Logging | Production-grade logging across all modules |
TruthLens employs a sequential multi-agent graph containing 7 specialized agents:
📥 Input → 🔍 ClaimExtractor → 🌐 EvidenceHunter → ✅ FactChecker → ⚖️ BiasAnalyzer → 📊 Verdict → 📝 Report
🔍 Verification Hub — Premium Dark Glassmorphism UI
📊 TruthLens Registry — System Metrics & Cache
🔌 Developer API — Integration Documentation
| Day | Concept | Implementation |
|---|---|---|
| 1 | Foundational Models | Gemini 2.5 Flash for reasoning, OCR, and evidence synthesis |
| 2 | Agents & Tools | 7 ADK agents with custom Python tools + MCP server (mcp_server.py) |
| 3 | Multi-Agent Systems | Sequential Workflow graph with structured inter-agent communication |
| 4 | Agent Quality | 24-test evaluation suite, golden dataset, structured logging |
| 5 | Deployment | HuggingFace Space + Docker + Cloud Run ready |
- Python 3.10+
- A Google Gemini API Key (Get one here)
# Clone the repository
git clone https://github.com/Rohith-Shimori/TruthLens-AI-Agent.git
cd TruthLens-AI-Agent
# Create virtual environment
python -m venv .venv
source .venv/bin/activate # Windows: .venv\Scripts\activate
# Install dependencies
pip install -r requirements.txt
# Configure API key
cp .env.example .env
# Edit .env and add your GOOGLE_API_KEY# Start the web UI
python app.py
# Open http://127.0.0.1:7860
# Run the evaluation suite
python -m tests.eval_suite --report
# Start the MCP server
python mcp_server.pydocker build -t truthlens .
docker run -p 7860:7860 --env-file .env truthlens======================================================================
🛡️ TruthLens Evaluation Suite — Golden Dataset Testing
======================================================================
📋 Security Manager: 6/6 ✅
📋 Credibility Scorer: 7/7 ✅
📋 Bias Analyzer: 4/4 ✅
📋 Memory Cache: 4/4 ✅
📋 SSRF Protection: 3/3 ✅
──────────────────────────────
📊 RESULTS: 24/24 tests passed (100.0% accuracy)
======================================================================
TruthLens-AI-Agent/
├── app.py # Main Gradio application (800+ lines)
├── mcp_server.py # FastMCP server for tool interoperability
├── requirements.txt # Pinned dependencies
├── Dockerfile # Multi-stage Docker build
├── watchdog.py # HA deployment watchdog
├── src/
│ ├── pipeline.py # ADK Workflow with 7 agents
│ ├── retrieval.py # Web scraping, Wikipedia, Google Grounding
│ ├── inference.py # Model config, retry policies, rate limiting
│ ├── utils.py # SecurityManager, MemoryManager, BiasAnalyzer
│ └── ui.py # Premium CSS, verdict cards, HTML templates
├── tests/
│ ├── eval_suite.py # 24-test evaluation suite with golden dataset
│ └── eval_report.json # Latest evaluation results
├── data/
│ └── sample_claims.json # 8 curated sample claims
├── docs/
│ └── kaggle_writeup.md # Competition writeup
└── assets/
├── logo.png
├── architecture.png
└── screenshots/
├── home.png
├── registry.png
└── developer_api.png
- Gemini Free-Tier Quota: High load may trigger
RESOURCE_EXHAUSTED(429). Handled via ADKRetryConfigwith exponential backoff. - Visual Language Context: OCR works best on English text screenshots. Highly distorted or handwritten regional text may reduce accuracy.
- Real-Time Data Lag: Breaking news within the last few minutes may have slight retrieval delay.
- 🌍 Multi-Lingual Verification — Translation agents for regional language claims
- 🎥 Audio & Video Transcription — Whisper-based transcription for TikTok/Reels/Shorts
- 🔄 Self-Correction Loop — LoopAgent for automated report quality validation
- 👥 RLHF Integration — Professional fact-checker feedback to improve agent prompts
© 2026 TruthLens • Kaggle × Google AI Agents Capstone
Built with ❤️ using Google ADK 2.0 & Gemini 2.5 Flash