📊 LLM–Powered Data Analytics Agent

A Streamlit-ready, LangChain-powered agent that enables natural language analytics on structured sales data using large language models (LLMs), machine learning, and statistical tooling.

---

🚀 Project Features

• 💬 Ask natural language questions about your sales data
• 📈 Forecast future sales using XGBoost
• ⚠️ Detect anomalies using Z-score method
• 🌍 Breakdown regional sales performance
• 💰 Analyze profitability across product lines
• 📊 Decompose time series into trend/seasonality
• 📝 Generate natural language sales reports

---

🗂️ Project Structure

llm_data_analytics_agent/
├── llm_powered_data_analytics_agent.py   # Main logic (LangChain + Streamlit + Tools)
├── cleaned_sales_data.csv                # Cleaned dataset used by the agent
├── memory.json                           # Memory file for conversation history
├── README.md                             # Project documentation

---

📦 Requirements

Install all dependencies in a virtual environment or Google Colab:

pip install -U openai langchain streamlit pandas numpy matplotlib seaborn plotly
pip install -U langchain-community langchain-openai langchain-experimental
pip install -U kaggle kagglehub pyngrok xgboost statsmodels mlxtend

---

🔑 Setup

1. Obtain Required Keys:

   • OpenAI API Key
   • Kaggle API Key (kaggle.json)

2. Environment Setup:

export OPENAI_API_KEY=your_openai_key
export KAGGLE_USERNAME=your_kaggle_username
export KAGGLE_KEY=your_kaggle_key

3. Kaggle Dataset Download (Optional):

kaggle datasets download -d kyanyoga/sample-sales-data
unzip sample-sales-data.zip

---

🧠 How It Works

The system uses the LangChain create_pandas_dataframe_agent with extra tools:

• SalesForecastingToolXGBoost
• AnomalyDetectionTool
• RegionSalesBreakdownTool
• ProfitabilityAnalysisTool
• TimeSeriesDecompositionTool
• ReportGeneratorTool

Each tool is wrapped with wrap_tool() to make it compatible with single-string input passed by the LangChain agent.

---

▶️ Running the Agent

✅ Option 1: Run in Google Colab

• Open llm_powered_data_analytics_agent.py in Colab
• Make sure the required userdata.get() entries are present
• Run cells sequentially
• Use dc_agent.run("Your question") to interact

✅ Option 2: Run with Streamlit

You can convert the script into a Streamlit app with a UI:

streamlit run llm_powered_data_analytics_agent.py

Update the script to support Streamlit input/output rendering.

---

💡 Example Queries

• "Forecast sales for the Motorcycles product line."
• "Detect any anomalies in the sales data."
• "What is the sales breakdown for APAC?"
• "Generate a summary report of performance."

---

🧩 Challenges Solved

• Handling single string input with wrap_tool()
• LLM tool selection using ZERO_SHOT_REACT_DESCRIPTION
• Formatting mixed outputs (tables, plots, markdown)
• Integrating memory (ConversationBufferMemory) across runs

---

📎 License

MIT License © 2025 — Anish Nilesh Rane

---