This project leverages the powerful Agent functionality of Claude Code to create a Multi-Agent System that simulates real data science team collaboration workflows. Through the collaborative work of AI Agents with different professional roles, it provides complete end-to-end data science solutions from raw data to final analysis reports.
- Simulate Real Teams: Recreate role division and collaboration patterns in data science projects
- Automated Workflow: Fully automated pipeline from data cleaning to model building
- Standardized Outputs: Generate professional-grade analysis reports, visualizations, and executable Notebooks
- Best Practices: Follow industry-standard project structure and code standards
| Role | Agent Name | Main Responsibilities | Input | Output |
|---|---|---|---|---|
| 🎯 Project Director | project-orchestrator |
Project planning, task scheduling, process management | User requirements | Final deliverables |
| 🔧 Data Engineer | data-engineer |
Data cleaning, preprocessing, feature engineering | Raw data | Clean datasets |
| 📊 Data Analyst | data-analyst-eda |
Exploratory data analysis, visualization | Processed data | EDA reports and charts |
| 🧠 ML Scientist | ml-model-scientist |
Model building, parameter tuning, performance evaluation | Analysis results | Predictive models and evaluations |
| 📝 Technical Writer | technical-writer-qa |
Report integration, quality assurance | All outputs | Final reports and Notebooks |
Our Multi-Agent system adopts a Sequential Pipeline architecture:
User → project-orchestrator
- Receive raw data and analysis requirements
- Develop project plan and timeline
- Initialize project structure
project-orchestrator → data-engineer → project-orchestrator
- Data quality checking and cleaning
- Missing value handling and outlier detection
- Feature engineering and data transformation
project-orchestrator → data-analyst-eda → project-orchestrator
- Descriptive statistical analysis
- Data visualization and trend discovery
- Correlation analysis and hypothesis testing
project-orchestrator → ml-model-scientist → project-orchestrator
- Model selection and training
- Hyperparameter tuning
- Model evaluation and validation
project-orchestrator → technical-writer-qa → project-orchestrator
- Technical documentation writing
- Quality assurance and review
- Final report generation
project-orchestrator → User
- Deliver complete analysis results
- Provide implementation recommendations and future planning
Each analysis project follows a standardized directory structure:
📁 project_[project_name]/
│
├── 📄 README.md # Project documentation
│
├── 📁 00_data/ # Data storage area
│ ├── 📁 raw/ # Raw data (immutable)
│ ├── 📁 processed/ # Cleaned data
│ └── 📁 intermediate/ # Intermediate processing files
│
├── 📁 01_notebooks/ # Jupyter Notebook workspace
│ ├── 📁 exploratory/ # Exploratory analysis drafts
│ ├── 📁 modeling/ # Model development experiments
│ └── 📁 final/ # Final executable Notebooks
│
├── 📁 02_src/ # Python source code
│ ├── 📄 data_processing.py # Data processing functions
│ ├── 📄 visualization.py # Visualization tools
│ └── 📄 modeling.py # Model training functions
│
├── 📁 03_reports/ # Reports and charts
│ ├── 📁 figures/ # Chart files
│ └── 📄 final_report.md # Final Markdown report
│
└── 📁 04_models/ # Trained models
└── 📄 [model_name].pkl
We use the classic Titanic dataset as an example to demonstrate the complete analysis workflow:
- Data Records: 891 training records
- Features: 12 original features
- Target Variable: Passenger survival rate (0/1)
- Best Model: Random Forest
- Accuracy: 85.5%
- F1 Score: 82.3%
- Python 3.10+
- Claude Code CLI
- Required packages: pandas, numpy, scikit-learn, matplotlib, seaborn
- Clone Project
git clone https://github.com/your-username/claudeCodeDataScienceTeam.git
cd claudeCodeDataScienceTeam- Setup Python Environment
# Use specified Python environment
C:/Users/hunghsun/AppData/Local/anaconda3/envs/py10/python.exe -m pip install -r requirements.txt- Launch Claude Code
claude- Prepare Data: Place raw data in
00_data/raw/directory - Start Analysis: Use
project-orchestratorAgent to begin new project - Monitor Progress: Observe the work results of each phase's Agent
- Review Results: Check
03_reports/final_report.mdfor complete analysis report
- Each Agent specializes in specific domains, ensuring professional quality
- Automated task handover and result validation
- Unified code style and document formats
- Reproducible analysis workflows and results
- Modular design, easy to add new Agent roles
- Support for different types of data science projects
- Built-in code checking and testing mechanisms
- Technical document review and quality control
- Project Independence: Each analysis project is completely independent, avoiding cross-contamination
- Separation of Concerns: Data, code, and reports are stored separately by category
- Raw Data Immutability: Ensure traceability and reproducibility of analysis
- Process-Oriented: Directory structure reflects the main data processing workflow
- Python Environment: Use Anaconda to manage package dependencies
- Visualization: Support Chinese fonts, charts directly saved as files
- Document Format: Markdown reports paired with high-quality PNG charts
- Version Control: Git manages code, model files stored independently


