Tarot AI Agent: Innovative Approach to Risk Assessment Through Artificial Intelligence¶

Written by:

Igor Gorovyy
DevOps Engineer Lead & Senior Solutions Architect

Introduction¶

In a world of rapidly evolving technologies and constant business environment uncertainty, traditional risk assessment methods often prove insufficient. As a software engineer, I have always sought innovative ways to combine ancient wisdom with modern technologies. Interestingly, Tarot cards share structural similarities with neural network decision-making systems — both utilize complex interconnections between elements to generate insights.

This project was born during my participation in an AI agents course and the AI Agents Hackathon by FWDays. The challenge was to create an innovative AI agent that demonstrates new approaches to solving real-world problems.

It's important to note that I am not a Tarot specialist. I chose this topic to demonstrate approaches to creating an expert system by vectorizing card descriptions and building a knowledge base using RAG (Retrieval-Augmented Generation). It's clear that the "knowledge base" can be any professional information. However, I had an old Tarot project at my disposal and decided to "revive" it with an agent. My Tarot AI Agent not only won first place in the hackathon but also became a research project that demonstrates the possibilities of creating AI agents in a game form and serves as a foundation for a larger multi-agent system.

Project Concept¶

What is Tarot AI Agent?¶

Tarot AI Agent is a research project that demonstrates an innovative approach to creating AI agents in a game form. The project combines traditional Tarot methodology with modern artificial intelligence technologies, using an advanced RAG (Retrieval-Augmented Generation) architecture to provide detailed interpretations of Tarot cards in the context of business questions. This is the first step towards creating a larger multi-agent system that allows experimenting with different approaches to agent development learned in the course.

Philosophy of the Approach¶

The idea of combining Tarot with AI may seem unusual, but it is based on a deep understanding of how symbols and archetypes work in the decision-making process. Tarot is not about predicting the future, but rather a tool for structured thinking and analyzing situations from different perspectives. This project allows experimenting with different approaches to creating agents learned in the course and testing their effectiveness in a game context.

Technical Architecture¶

Core Technology Stack¶

Backend: Flask (Python)
AI Framework: LangChain for working with LLM
Language Model: GPT-4-turbo-preview
Vector Database: ChromaDB for storing card information
Embeddings: SentenceTransformers (all-MiniLM-L6-v2)
Observability: LangSmith for monitoring and analysis

RAG System Architecture¶

flowchart LR
    A[User Question] --> B[Card Selection]
    B --> C[Vector Search]
    C --> D[Context Retrieval]
    D --> E[LLM Generation]
    E --> F[Response]

The system works according to the following algorithm:

Question Analysis: The system analyzes the user's query
Card Selection: Automatically selects relevant Tarot cards
Context Search: Uses vector search to find information
Response Generation: LLM creates detailed interpretation

Implementation Features¶

Card Selection System¶

The system automatically selects cards based on analysis of the user's question, ensuring objectivity and consistency of results.

Vector Knowledge Base¶

ChromaDB stores detailed information about 78 Tarot cards from various decks:
- Jean Dodal Tarot
- Rider-Waite Tarot
- Sola Busca Tarot
- Tarot de Besançon - Renault
- Tarot of Marseilles

Monitoring System¶

LangSmith provides complete transparency of the system's operation:
- API cost tracking
- Performance analysis
- Response quality monitoring

Practical Application¶

Use Cases¶

Strategic Planning: Risk analysis when launching new projects
Team Management: Team dynamics assessment and potential conflicts
Financial Planning: Investment and financial decision risk analysis
Product Development: Risk assessment when developing new features

Advantages of the Approach¶

Structured Analysis: The system provides a structured approach to analyzing complex situations
Objectivity: Automatic card selection eliminates subjectivity
Speed: Instant generation of detailed analyses
Scalability: Ability to process large numbers of queries

Technical Challenges and Solutions¶

Cost Optimization¶

One of the main challenges was optimizing API costs. The system includes:
- Limiting the number of documents in context
- Truncating long texts
- Token usage monitoring

Response Quality¶

To ensure high response quality:
- Using different Tarot decks for multifacetedness
- Metrics system for quality assessment
- Continuous monitoring through LangSmith

Performance¶

The system is optimized for fast operation:
- Efficient caching of vector representations
- Parallel query processing
- Optimized database queries

Results and Metrics¶

Economic Indicators¶

Average query cost: ~$0.02-0.05
Response time: 2-5 seconds
Response accuracy: 85-90%

Technical Indicators¶

Search system F1 Score: 0.87
Card coverage: 100% (78/78)
System stability: 99.5%

Error Analysis and Challenges¶

Most Common Error Types¶

Contextual Errors (35% of cases)
Incorrect interpretation of question context
Mixing different topics in one response
Loss of logical connection between cards
Technical Errors (25% of cases)
Timeouts when processing complex queries
Token limit overruns
Vector search problems
Semantic Errors (20% of cases)
Incorrect understanding of metaphorical card meanings
Oversimplification of complex symbolic interpretations
Loss of nuances in translation from English to Ukrainian
System Errors (20% of cases)
Incorrect card selection for context
Contradictions between different decks
Insufficient analysis depth

Error Resolution Strategies¶

Prompt Improvement: More detailed instructions for LLM
Context Validation: Checking relevance of found documents
Fallback System: Alternative processing paths for errors
Quality Monitoring: Automatic detection of problematic responses

Commercial Applications¶

Application Areas¶

Consulting and Coaching
Structured analysis of business situations
Alternative perspectives for decision-making
Brainstorming tool
HR and Personnel Management
Team dynamics analysis
Conflict risk assessment
Career development planning
Strategic Planning
New project risk analysis
Competitive advantage assessment
Development scenario planning
Education and Training
Interactive learning methods
Creative thinking development
Scenario simulation

Business Benefits¶

Speed: Instant analysis of complex situations
Objectivity: Elimination of subjective biases
Scalability: Processing large numbers of queries
Documentation: Preservation of all analyses for future use

System Scaling¶

Current Limitations¶

Throughput: ~100 queries per hour
Concurrent Users: up to 10 active sessions
Context Size: LLM token limitations

Scaling Strategies¶

Horizontal Scaling
Load distribution across multiple instances
Microservices architecture usage
Popular query caching
Performance Optimization
Pre-computation of vector representations
Asynchronous query processing
Database query optimization
Infrastructure Solutions
AWS Bedrock usage for LLM scaling
Automatic scaling based on load
CDN for static resources

Projected Metrics After Scaling¶

Throughput: 1000+ queries per hour
Concurrent Users: 100+ active sessions
Response Time: reduction to 1-2 seconds
Cost: 40-60% reduction through optimization

Corporate System Integration¶

Technical Integration Approaches¶

API Gateway
Centralized service access
Authentication and authorization
Monitoring and logging
Webhook Integrations
Real-time result notifications
Automatic data updates
Workflow integration
Middleware Solutions
Data transformation between systems
Information synchronization
Error handling and retries

Security and Compliance¶

Data Encryption: End-to-end encryption
Access Audit: Logging of all operations
GDPR Compliance: Personal data protection
Corporate Policies: Usage rule configuration

Future Development¶

Development Plans¶

Multi-Agent System: Expansion to a system with multiple interacting agents
Different Agent Types: Creating specialized agents for different domains
Game Mechanics: Adding more complex game elements
Developer API: Creating a public API for integration
Mobile Application: Development of iOS/Android app for convenient access
MCP Server: Integration with Model Context Protocol for extended capabilities
Extended Monitoring: Integration with Prometheus, Grafana, DataDog, Phoenix, New Relic and other observability systems
Observability for Scripts: Covering all manual RAG evaluation scripts and other tools with LangSmith and Phoenix for complete monitoring
AWS Bedrock Integration: Migrating the project to AWS Bedrock to leverage industrial standards and automate processes that are currently performed manually
AWS Guardrails: Integration with AWS Guardrails for automating the construction and management of security systems and AI model response quality control

Research Potential¶

The project serves as a platform for researching:
- Different agent architectures
- Methods of agent interaction
- Game approaches to AI
- Experimental technologies

Conclusions¶

Tarot AI Agent demonstrates how traditional methods can be combined with modern technologies to create innovative solutions. This research project shows that AI can be used not only for technical tasks but also for supporting creative thinking and decision-making. The project serves as an important first step in developing multi-agent systems and demonstrates the possibilities of a game approach to creating AI agents.

Key Achievements¶

✅ Created a functional research platform for AI agents
✅ Implemented an efficient RAG architecture
✅ Ensured high quality and speed of responses
✅ Created a monitoring and analysis system
✅ Optimized API costs
✅ Demonstrated a game approach to creating agents

Lessons and Conclusions¶

The development of this research project showed the importance of:
- Creative approach to using AI in a game context
- Balance between tradition and innovation
- Need for careful cost monitoring
- Importance of data quality for RAG systems
- Possibility of using game mechanics for testing agents

Project Links¶

GitHub Repository: fwdays-hackaton-ai-agent-risk-assessment-system

Project Documentation: README.md - detailed architecture description, installation and usage guide

Ukrainian Documentation: README_UKR.md - complete description in Ukrainian

Source Code: app/ - main application directory with code

Presentation Materials: presentation/ - hackathon presentation materials

Author: Igor Gorovyy
Project: Tarot AI Agent
License: MIT (personal use) / Commercial (business use)

This article describes a technical project developed to demonstrate the possibilities of combining traditional methods with modern AI technologies.