SLENTH - AML Monitoring System

Agentic AI for Real-Time AML Monitoring and Document Corroboration

Julius Baer Hackathon Solution

🎉 Implementation Status: ~60% Complete (Infrastructure Ready)

---

📊 Quick Status

Component	Status	Progress
Services Layer	✅ Complete	100%
API Endpoints	✅ Complete	100%
Workflows	✅ Complete	100%
Agent Skeletons	✅ Complete	100%
Agent Logic	⏳ In Progress	10%
Crawlers	⏳ Pending	0%

📋 See IMPLEMENTATION_STATUS.md for detailed breakdown

---

🎯 Project Overview

SLENTH is a comprehensive Anti-Money Laundering (AML) compliance system that uses agentic AI workflows powered by LangGraph to provide:

Part 1: Real-Time AML Transaction Monitoring

• Automated ingestion of external regulatory circulars (HKMA, MAS, FINMA)
• API-driven internal rules management with versioning
• Real-time transaction compliance evaluation using RAG
• Multi-role alerts (Front Office, Compliance, Legal)
• Automated remediation workflows

Part 2: Document & Image Corroboration

• Multi-format document processing (PDF, images, text)
• OCR extraction and format validation
• Image forensics (tampering detection, EXIF analysis, AI-generated detection)
• Background checks via World-Check One API
• Cross-reference with transaction history

---

🏗️ Architecture

Processing Modes

• Part 1: Async processing via Celery + Redis queue (high throughput)
• Part 2: Synchronous processing (immediate results)

Infrastructure

• Database: Cloud PostgreSQL (structured data)
• Vector DB: Qdrant (regulatory rules, RAG)
• Queue: Redis (Part 1 only)
• LLM: OpenAI GPT-4 / Anthropic Claude
• Embeddings: text-embedding-3-large

---

📁 Project Structure

See full structure in the project. Key components:

• app/ - FastAPI application and API endpoints
• agents/ - LangGraph agents (Part 1: 13 agents, Part 2: 10 agents)
• workflows/ - LangGraph workflow orchestration
• services/ - Business logic (vector DB, LLM, embeddings, alerts)
• db/ - Database models and connections
• crawlers/ - Regulatory website scrapers
• worker/ - Celery workers (Part 1 only)
• scripts/ - Utility scripts and simulators

---

🚀 Quick Start

Prerequisites

• Python 3.10+
• Cloud PostgreSQL database
• Redis (local via Homebrew)
• Docker (for Qdrant)
• Tesseract OCR
• OpenAI API key

Installation

# Clone and setup
git clone https://github.com/clarud/slenth.git
cd slenth

# Create virtual environment
python -m venv venv
source venv/bin/activate

# Install dependencies
pip install -r requirements.txt
playwright install

# Install Tesseract OCR (macOS)
brew install tesseract

# Configure environment
cp .env.example .env
# Edit .env with your credentials

# Start infrastructure
brew services start redis
docker run -d -p 6333:6333 -p 6334:6334 \
  -v $(pwd)/qdrant_storage:/qdrant/storage \
  --name slenth-qdrant qdrant/qdrant

# Initialize databases
python scripts/init_db.py
python scripts/init_vector_db.py
python scripts/load_internal_rules.py

# Run application
# Terminal 1:
uvicorn app.main:app --reload --host 127.0.0.1 --port 8000

# Terminal 2 (Part 1 only):
celery -A worker.celery_app worker -l info -Q default -c 4

Access Points

• API: http://localhost:8000
• API Docs: http://localhost:8000/docs
• Health: http://localhost:8000/health

---

📡 API Examples

Part 1: Process Transaction (Async)

curl -X POST http://localhost:8000/transactions \
  -H "Content-Type: application/json" \
  -d '{
    "transaction_id": "TXN001",
    "amount": 50000,
    "currency": "USD",
    "customer_id": "CUST123",
    "originator_country": "HK",
    "beneficiary_country": "SG"
  }'

Part 2: Upload Document (Synchronous)

curl -X POST http://localhost:8000/documents/upload \
  -F "file=@data/Swiss_Home_Purchase_Agreement_Scanned_Noise_forparticipants.pdf" \
  -F "document_type=purchase_agreement"

---

📝 Implementation Status

✅ Completed

• Project structure
• Configuration management
• Database models (14 tables)
• Main FastAPI app
• Health check endpoints
• Requirements and dependencies
• Documentation

🚧 To Be Implemented

Based on AGENTIC_AML_WORKFLOW_PLAN.md, you need to implement:

1. API Endpoints (app/api/)

   • transactions.py, documents.py, internal_rules.py, alerts.py, cases.py

2. Agents (agents/)

   • Part 1: 13 agents for transaction monitoring
   • Part 2: 10 agents for document corroboration

3. Workflows (workflows/)

   • LangGraph workflows for Part 1 and Part 2

4. Services (services/)

   • Vector DB, embeddings, LLM, alerts, World-Check One

5. Crawlers (crawlers/)

   • HKMA, MAS, FINMA regulatory scrapers

6. Worker (worker/)

   • Celery configuration and tasks

7. Scripts (scripts/)

   • Database init, transaction simulator

---

📚 Documentation

• Full Implementation Plan: AGENTIC_AML_WORKFLOW_PLAN.md
• Problem Statement: README copy.md
• API Docs: http://localhost:8000/docs (when running)

---

🔧 Development

Follow implementation order:

1. Services layer (vector DB, LLM, embeddings)
2. Workflow foundations
3. Part 1 agents and workflow
4. Part 2 agents and workflow
5. API endpoints
6. Crawlers and cron jobs
7. Testing

---

Built for Julius Baer Hackathon 2025 🚀