Multimodal Agentic Business Intelligence Analyst (ABIA)

An autonomous AI agent crew that transforms raw database metrics into a comprehensive, multimodal business report—complete with data visualizations, external market research, and executive synthesis—in under 60 seconds.

---

🏗️ System Architecture & Workflow

The system utilizes a sequential multi-agent orchestration pipeline. It is governed by a custom-engineered LLMFailoverRouter that prioritizes extreme execution speed while guaranteeing high availability.

flowchart TD
    subgraph DataLayer ["Data Layer"]
        DB[("DuckDB\nSales Data")]
    end

    subgraph FailoverRouter ["LLM Failover Router"]
        Router{"API Gateway"}
        Groq["Groq Llama 3\nPrimary - Fast"]
        Gem["Google Gemini\nFallback - Smart"]
        Router -->|"Success"| Groq
        Router -->|"Rate Limited / 429"| Gem
    end

    subgraph CrewAI ["CrewAI Orchestration"]
        A1["Data Analyst Agent"]
        A2["Market Researcher Agent"]
        A3["Reporting Analyst Agent"]
        
        DB -->|"SQL Tools"| A1
        A1 -->|"Chart + Summary"| A2
        A2 -->|"Web Search Tools"| A3
    end

    Router -.->|"Powers"| A1
    Router -.->|"Powers"| A2
    Router -.->|"Powers"| A3
    A3 -->|"Compiles"| Output(["Multimodal PDF Report"])

    style Router fill:#f9a826,stroke:#333,stroke-width:2px,color:#000
    style Output fill:#FF4B4B,stroke:#333,stroke-width:2px,color:#fff

Loading

---

⚙️ Key Engineering Features

• Robust LLM Failover Strategy: The core of the pipeline's reliability. A custom LLMFailoverRouter assigns a high-speed Groq model (llama-3.1-8b-instant) to all agents for a sub-60-second runtime. If the Groq API hits a rate limit (HTTP 429), the router automatically and silently fails over to a backup model (gemini-1.5-flash) without crashing the execution state.
• Autonomous Agent Crew: Employs specialized AI personas (Data Analyst, Market Researcher, Reporting Analyst) built on CrewAI to collaborate, delegate, and achieve complex analytical goals sequentially.
• Custom Tool Engineering: Agents operate autonomously using a suite of custom-built tools for localized SQL execution (DuckDB), file I/O operations, programmatic data visualization (Seaborn/Matplotlib), and real-time web crawling (Tavily).
• Multimodal Delivery: The final pipeline deliverable is an automated PDF report that programmatically embeds generated data visualization charts (.png) alongside the synthesized executive text and full agent execution logs.

---

💻 Technology Stack

• Agentic Framework: CrewAI, LangChain
• LLM Providers: Groq (Primary), Google Gemini (Failover)
• Data Backend: DuckDB, Pandas
• Visualization: Matplotlib, Seaborn
• Infrastructure & Tools: Tavily API (Search), FPDF2 (Reporting)

---

🚀 Reproducibility & Setup

1. Repository Setup

Clone the repository and install the required dependencies (Python 3.10+ recommended):

git clone [https://github.com/AD1007/Multimodal-Agentic-Business-Intelligence-Analyst.git](https://github.com/AD1007/Multimodal-Agentic-Business-Intelligence-Analyst.git)
cd Multimodal-Agentic-Business-Intelligence-Analyst

pip install -r requirements.txt

2. Environment Configuration

Create a .env file in the root directory to authenticate the LLM router and search tools:

GROQ_API_KEY="your_groq_key_here"
GOOGLE_API_KEY="your_gemini_key_here"
TAVILY_API_KEY="your_tavily_key_here"

3. Pipeline Execution

Execute the master orchestration script. The system will auto-generate the DuckDB instance, trigger the agent crew, and compile the final artifacts in real-time.

python RunPipeline.py

The final multimodal report will be saved as ABIA_Report.pdf in your project root.

---

📂 Project Structure

.
├── demo/
│   ├── ABIA_Report.pdf          # Pre-compiled report for live demo
│   └── index.html               # GitHub Pages entry point
├── data/
│   └── gadget_grove.db          # Auto-generated DuckDB instance
├── outputs/
│   └── monthly_revenue_plot.png # Agent-generated visualization
│
├── agents_and_tasks.py          # CrewAI roles and task definitions
├── data_setup.py                # Database synthesis and population
├── llm_setup.py                 # LLM definitions and Failover Router logic
├── tool_development.py          # Custom agent tools (SQL, Search, Viz)
├── RunPipeline.py               # Main execution and PDF compilation
│
├── .env                         # API Credentials
├── README.md                    # System documentation
└── requirements.txt             # Dependency constraints

License

This project is licensed under the MIT License.