TradeScreenAI

AI-Powered Multi-List Sanctions Screening with Cyrillic Transliteration

🔗 Live Demo: tradescreenai.com

An AI-driven sanctions screening system that screens vendors against 4 international sanctions lists (45,296+ entities) with specialized Cyrillic-to-Latin transliteration and GPT-4o deep analysis. This prototype was developed as the final project for the Information Technology Systems course. All work was performed individually as part of academic requirements.

Author: Tatiana Podobivskaia
Affiliation: Atlantis University, Miami, FL
Program: MBA in Business Intelligence & Data Analytics

---

Sanctions Lists

List	Source	Entities	Update
OFAC SDN	U.S. Treasury	18,714	Daily
EU Consolidated	European Union	5,819	Daily
UN Security Council	United Nations	1,002	Daily
UK OFSI	UK Government	19,761	Daily
Total	4 Lists	45,296

Key Features

• Multi-list screening — simultaneous screening across OFAC SDN, EU, UN, and UK OFSI sanctions lists
• Cyrillic transliteration engine — generates 2–5 Latin-script variants per name using ISO 9, ICAO, BGN/PCGN, and informal systems
• Two-pass architecture — Pattern Matching (instant, in-browser) + AI Deep Analysis (Azure OpenAI GPT-4o)
• Multi-algorithm fuzzy matching — composite scoring using token sort, token set, partial, and standard ratio
• Alias matching — checks entity aliases from all 4 lists
• PDF compliance reports — auto-generated with executive summary, risk breakdown, and audit trail
• Risk-based routing — automated APPROVE / FLAG / BLOCK with configurable thresholds

Problem

• Existing enterprise compliance tools are often cost-prohibitive for academic research environments, motivating this prototype approach.
• Cyrillic transliteration creates detection gaps — "Щербаков" can appear as Shcherbakov, Scherbakov, Shtcherbakov, or Shherbakov
• Single-list screening is insufficient — sanctioned entities may appear on EU/UN/UK lists but not OFAC

Cyrillic Transliteration

Variant	System
Shcherbakov	ISO 9
Scherbakov	Simplified / Informal
Shtcherbakov	Passport-style
Schtscherbakov	German-influenced

Standard tools treat these as different entities. TradeScreenAI generates all variants and matches against all 4 lists simultaneously.

Architecture

Trade Document (CSV / Single Vendor)
        │
        ▼
   Entity Extraction
        │
        ▼
Cyrillic Transliteration ──► Generate 2–5 Latin variants
        │
        ▼
Multi-List Screening ──► OFAC SDN + EU + UN + UK OFSI (45,296 entities + aliases)
        │
        ▼
   Composite Risk Score
        │
        ├── LOW  (< 50)  ──► Auto-approve, Log
        ├── MED  (50–84) ──► Flag, Notify compliance officer
        └── HIGH (≥ 85)  ──► Block, Escalate
        │
        ▼
   AI Deep Analysis (Azure OpenAI GPT-4o)
        │
        ▼
   PDF Compliance Report

Quick Start

Installation

git clone https://github.com/tatianapodobivskaia-del/trade-document-intelligence.git
cd trade-document-intelligence
pip install -r requirements.txt

Run Multi-List Screening Demo

python sdn_matcher.py

This will:

• Download all 4 sanctions lists (OFAC SDN, EU, UN, UK OFSI)
• Demonstrate Cyrillic transliteration variants
• Screen 7 sample vendors against 45,296 entities
• Export results to CSV and JSON

Screen Your Own Vendors

from sanctions_lists import load_all_sanctions_lists
from sdn_matcher import screen_vendor

data = load_all_sanctions_lists()
result = screen_vendor("Rosoboronexport Trading", data["entities"],
                       cyrillic_name="Рособоронэкспорт Трейдинг")
print(result)

Project Structure

trade-document-intelligence/
├── index.html              # Web interface (tradescreenai.com)
├── sanctions_lists.py      # Multi-list loader (OFAC + EU + UN + UK OFSI)
├── sdn_matcher.py          # Screening engine with transliteration
├── privacy.html            # Privacy Policy
├── terms.html              # Terms of Service
├── disclaimer.html         # Compliance Disclaimer
├── requirements.txt        # Python dependencies
├── sample_vendors.csv      # Test data
├── screening_results.csv   # Output (generated)
├── screening_results.json  # JSON export (generated)
└── README.md

Technology Stack

• Python — Screening engine and transliteration
• Microsoft Azure — Cloud infrastructure (Functions, OpenAI)
• Azure OpenAI GPT-4o — AI deep analysis for true/false positive classification
• Vercel — Web hosting and CDN
• OFAC, EU, UN, UK OFSI — Official government sanctions data

Risk Scoring

Composite score (0–100) from multiple algorithms:

Algorithm	Weight	Strength
Token Sort Ratio	30%	Handles word order differences
Token Set Ratio	30%	Handles extra/missing words
Partial Ratio	25%	Catches substring matches
Standard Ratio	15%	Overall string similarity

Weighted Composite: Name Match (75%) + Country Risk (10%) + Amount (5%) + Document Type (5%) + Cyrillic Bonus (5%)

Dependencies

requests
fuzzywuzzy
python-Levenshtein
transliterate
pandas

Disclaimer

TradeScreenAI is a research prototype developed as part of graduate-level coursework. It is a decision-support tool for informational and educational purposes only. It does not constitute legal or compliance advice. See disclaimer.

License

Copyright © 2026 Tatiana Podobivskaia. All rights reserved. This software is proprietary. Unauthorized copying, modification, or distribution is prohibited.

Contact

Tatiana Podobivskaia
Email: tatiana.podobivskaia@atlantisuniversity.edu
Website: tradescreenai.com
LinkedIn: linkedin.com/in/tatiana-podobivskaia