An AI agent that investigates a real business question against a real SQLite database — through a real MCP server — and writes an evidence-grounded report instead of a chatbot guess.
Eighth deliverable of a 90-day AI Engineering roadmap (Phase 1: Foundation, Week 8).

📑 Table of Contents

Overview
Features
Demo
Installation
How to Run
Example Question
Architecture
Why a Real Database + Real MCP Server
Folder Structure
Future Improvements
Roadmap Context
Author
License

📖 Overview

AI Business Agent (UI codename Ledger) takes a plain-English business
question — "Why did our revenue decrease this month?" — and actually
investigates it: classifies the question, plans which business data it
needs, queries a real SQLite database (optionally through a real MCP
server), computes grounded findings in plain Python, optionally gathers
external research, and writes a structured report. If the data doesn't
support a conclusion, it says so instead of guessing.

This is Repo 8 of 10+ in a structured 90-day AI Engineering roadmap.

✨ Features

🧭 Question Classification — maps a business question to a category (sales, customers, operations, finance, inventory, marketing, general) and the data it needs
🗃️ Real SQLite Business Database — customers, products, orders, sales tables, generated from realistic synthetic data (Pakistani regions: Punjab, Sindh, KPK, Balochistan)
🔌 Real MCP Server — mcp_server.py exposes 7 business-data tools over the actual Model Context Protocol (official mcp SDK), connectable from Claude Desktop, Claude Code, or any MCP client — not just this app
🧠 Deterministic Planning — planner.py extracts the month/region/product mentioned in the question and selects the right tools, with no LLM call needed for that step
📊 Grounded Analysis — analyzer.py computes percentage changes and distinguishes a price-driven decline from a demand-driven one, directly from the numbers — not from LLM guesswork
🌐 Optional External Research — browser_tools.py (Tavily search), used only when the question asks for external factors, with an honest "not configured" fallback instead of silent failure
📁 Sandboxed File Access — file_tools.py can only read from data/, docs/, notes/; path traversal is structurally blocked
🚫 No Hallucinated Numbers — the report-writing LLM only works from evidence it's explicitly handed; "Insufficient evidence to determine the cause" is a real, tested fallback
🎨 Evidence-Transparent UI — every report comes with an expandable "agent's work" panel showing the plan, tool calls, and raw findings
🧪 Evaluated against 15+ real business questions — sales, customers, operations, research, multi-step, and deliberately impossible/insufficient-evidence questions

🎥 Demo

(Add a screenshot or short GIF/video here once available — see assets/screenshots/)

🛠️ Installation

# 1. Clone the repository
git clone https://github.com/Hamna-Munir/08-AI-Business-Agent.git
cd 08-AI-Business-Agent

# 2. Create a virtual environment
python -m venv venv
source venv/bin/activate      # On Windows: venv\Scripts\activate

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

# 4. Configure environment variables
cp .env.example .env
# then add your Groq/LLM API key (OPENAI_API_KEY, starts with gsk_)
# TAVILY_API_KEY is optional — only needed for external-research questions

▶️ How to Run

# Build the synthetic business database (safe to re-run; no-ops if it exists)
python -m src.database

# Run the Streamlit app
streamlit run src/app.py
# or: python -m streamlit run src/app.py

# Run the MCP server standalone (for Claude Desktop/Code or the MCP Inspector)
python -m src.mcp_server

# Run the test suite (fully mocked LLM calls — no API key needed)
pytest

💬 Example Question

"Why did our revenue drop this month?"

The agent finds — for real, from the database, not scripted —:

UrbanTote Backpack revenue decreased 35.7% from 2026-05 (PKR 167,162) to
2026-06 (PKR 107,464). Order count changed +0.0% and units sold changed
-6.2% over the same period. Average unit price changed -31.2%. Because
order volume stayed roughly stable while average price dropped sharply,
the revenue decline is price-driven, not demand-driven.

🏗️ Architecture

Business Question
        ↓
   BUSINESS AGENT
        ↓
   Task Planning            (Day 50 classify + Day 55 plan)
        ↓
   Tool Selection
        ↓
┌───────────┼────────────┐
↓           ↓            ↓
MCP Tools  Database    Browser/File
(Day 52)   (Day 53)    (Day 54)
↓           ↓            ↓
└───────────┼────────────┘
        ↓
   Evidence/Data
        ↓
   Analysis               (Day 55-56 — grounded, numeric)
        ↓
   Business Recommendations
        ↓
   Structured Report

🔍 Why a Real Database + Real MCP Server

A toy in-memory dataset with no real pattern in it can't teach — or
demonstrate — anything about evidence-based reasoning, because every
question would have either an obvious or an arbitrary answer. The SQLite
database here has a deliberately engineered, non-obvious story in it (see
docs/week-08-summary.md), and mcp_server.py exposes the same business
tools over the actual MCP protocol rather than a bespoke function-calling
shim — so the same data is queryable from this app, from Claude Desktop, or
from any other MCP client, which is the actual point of MCP.

📂 Folder Structure

08-AI-Business-Agent/
│
├── README.md
├── LICENSE
├── .gitignore
├── requirements.txt
├── .env.example
│
├── src/
│   ├── __init__.py
│   ├── app.py
│   ├── agent.py
│   ├── planner.py
│   ├── business_tools.py
│   ├── mcp_server.py
│   ├── database.py
│   ├── browser_tools.py
│   ├── file_tools.py
│   ├── analyzer.py
│   ├── prompts.py
│   └── config.py
│
├── data/
│   ├── customers.csv
│   ├── products.csv
│   ├── orders.csv
│   ├── sales.csv
│   └── business.db
│
├── tests/
│   ├── __init__.py
│   ├── test_agent.py
│   ├── test_tools.py
│   ├── test_database.py
│   ├── test_mcp.py
│   └── test_analysis.py
│
├── evaluation/
│   └── evaluation.csv
│
├── docs/
│   └── week-08-summary.md
│
├── notes/
│   ├── day-50.md
│   ├── day-51.md
│   ├── day-52.md
│   ├── day-53.md
│   ├── day-54.md
│   ├── day-55.md
│   └── day-56.md
│
├── assets/
│   ├── banner.svg
│   └── screenshots/
│
└── journal.md

🚀 Future Improvements

LLM-based entity extraction in planner.py for messier real-world phrasing
Multi-turn "drill into this finding" follow-up mode
Point database.py at a real data warehouse/CSV export instead of synthetic data
Add a second external-research provider as a fallback to Tavily
Export reports as PDF/Word for sharing with stakeholders

🧭 Roadmap Context

This project is Week 8 of Phase 1 in a 90-day AI Engineering roadmap:

Phase	Focus	Days
Phase 1	Foundation — Personal Assistant → ... → Meeting Agent → Business Agent	1–30
Phase 2	Agent Engineering — RAG, LangGraph, MCP	31–60
Phase 3	Business AI Systems — Multi-Agent, Deployment	61–90

👩‍💻 Author

Hamna Munir
Software Engineering & AI/ML Student | Building deployable AI/ML projects

GitHub: @Hamna-Munir
Hugging Face: @Hamna27

📄 License

This project is licensed under the MIT License — see the LICENSE file for details.