SanctionGuard is an AI-powered compliance platform that replaces traditional "keyword matching" with a multi-agent reasoning system. Instead of simply flagging a name, it convenes a digital tribunal to debate the risk, drastically reducing false positives.
https://sanctionguard-aicompliancechecking-9d6mzvvkbyfmz8wuznjtyq.streamlit.app/
- Multi-Agent Tribunal:
- 👨⚖️ The Prosecutor (Llama 3.3): Analyzes name matches and country risks to argue for strict enforcement.
- 🛡️ The Defense (GPT-OSS 20B): Identifies data gaps, typos, and lack of biometric evidence to argue for false positives.
- ⚖️ The Judge (Gemini 2.5): Synthesizes both arguments into a final, explainable verdict (High/Low Risk).
- Tiered Architecture: Uses RapidFuzz (Tier 1) for speed and LLMs (Tier 2) for reasoning, optimizing cost and latency.
- Batch Screening: Upload a CSV to screen hundreds of entities automatically.
- Audit-Ready Reports: Generates downloadable PDF case files with full tribunal transcripts.
- Frontend: Streamlit
- AI Inference: Groq (Llama 3.3, GPT-OSS, Mixtral) & Google Gemini
- Data Processing: Pandas, RapidFuzz
- Reporting: FPDF
- Data Source: Consolidated OFAC & UN Sanctions List (JSON)
-
Clone the repository:
git clone [https://github.com/yourusername/sanctionguard-ai.git](https://github.com/yourusername/sanctionguard-ai.git) cd sanctionguard-ai -
Install dependencies:
pip install -r requirements.txt
-
Set up API Keys: Create a
.streamlit/secrets.tomlfile (DO NOT commit this to GitHub):GROQ_API_KEY = "gsk_..." GOOGLE_API_KEY = "AIza..."
-
Run the application:
streamlit run app.py
app.py: Main Streamlit dashboard and UI logic.tribunal.py: The core AI logic class handling the "Trial" and model interactions.evidence_manager.py: Utility to download and parse the latest sanctions lists (OFAC/UN).consolidated_sanctions.json: The local vector/search database.
This project is for educational and research purposes.