🤖 Agent Jarvis — Autonomous AI Database Agent

A fully autonomous AI Database Agent built from scratch using pure Python — no LangChain, no CrewAI, no drag-and-drop frameworks. Just raw agent architecture: a reasoning loop, custom tools, memory management, and a conversational web UI.

Built during a software development internship over 16 days.

---

🧠 How It Works

Most people build "AI agents" by chaining API calls or dragging nodes on n8n. This project builds the real thing — from the ground up.

User Message
     │
     ▼
 Reasoning Loop  ◄─────────────────────┐
     │                                  │
     ▼                                  │
Tool Selection (JSON Schema)            │
     │                                  │
     ▼                                  │
Tool Execution                          │
  ├── ask_database()  → MySQL           │
  ├── create_file()   → Local FS        │
  ├── read_file()     → Local FS        │
  └── get_time()      → System          │
     │                                  │
     ▼                                  │
Observe Result ────────────────────────►┘
     │
     ▼ (task complete)
 Final Response → Chainlit UI

The agent keeps looping — calling tools, observing results, deciding next steps — until the task is fully complete. No human in the loop.

---

🏗️ Project Structure

Database Agent/
│
├── .chainlit/
│   ├── translations/
│   └── config.toml
│
├── .files/
│   └── .env                  # API keys and DB credentials (not committed)
│
├── .gitignore
├── chainlit.md               # Chainlit welcome screen content
├── database_agent.py         # Core agent — reasoning loop, tools, memory, UI
├── schema_prompt.py          # MySQL schema definitions and SQL rulebook
└── README.md

---

⚙️ Features

• Inner Autonomous Loop — Agent calls tools back-to-back without human intervention until the task is done
• 6 Custom Tools with full JSON schemas registered via the OpenRouter tool-calling API:
  • ask_database — Translates natural language → SQL → executes on MySQL (sub-agent pattern)
  • create_file — Creates and writes local text files
  • read_file — Reads local file contents
  • get_time — Fetches current system timestamp
  • say_hello — Greeting function
  • get_agent_name — Returns agent identity
• Sub-Agent Pattern — ask_database() runs its own internal LLM call to generate SQL before touching MySQL
• Sliding Window Memory — Conversation history capped at 20 turns to stay within context limits
• Custom JSON Encoder — Handles MySQL Decimal and datetime types that standard json.dumps() cannot serialize
• Chainlit Web UI — Dark-themed conversational interface with real-time multi-turn support

---

🛠️ Tech Stack

Layer	Technology
Language	Python 3.10+
LLM Provider	OpenRouter (nvidia/nemotron / openrouter/free)
SQL Generation	Gemini API (google-genai)
Database	MySQL via pymysql
Web UI	Chainlit
Environment	python-dotenv

---

🚀 Getting Started

1. Clone the repository

git clone https://github.com/Jainil26/Ai-Database-Agent.git
cd Ai-Database-Agent

2. Install dependencies

pip install -r requirements.txt

3. Set up environment variables

Create a .env file inside the .files/ folder:

OPENROUTER_API_KEY=your_openrouter_api_key
DB_HOST=localhost
DB_USER=your_mysql_user
DB_PASSWORD=your_mysql_password
DB_NAME=your_database_name

4. Set up the MySQL database

Run your schema setup script in MySQL Workbench or via CLI. The agent expects these tables:

categories (category_id, name)
products (product_id, category_id, product_name, status)
product_variants (variant_id, product_id, sku, price)
customers (customer_id, first_name, last_name, email)
orders (order_id, customer_id, order_status, total_amount)

5. Run the agent

Web UI mode (Chainlit):

chainlit run database_agent.py

Then open http://localhost:8000 in your browser.

---

🐛 Real Bugs Solved

These are actual issues hit during development — not hypothetical:

Bug	Cause	Fix
Infinite loop	cursor.fetchall missing () — referenced, never called	Added parentheses to execute the function
JSON crash	MySQL returns Decimal objects for financial fields	Built DatabaseJsonEncoder class
Schema hallucination	LLM invented order_date column that didn't exist	Tightened SQL rulebook in schema_prompt.py
KeyError on history	Inconsistent dictionary keys across memory turns	Synchronized all keys across the loop
MySQL Error 1064	LLM returned plain English instead of SQL	Added HISTORY_REQUEST keyword routing to bypass DB

---

📚 What I Learned

Building from scratch means no framework hides the errors. Which means no framework hides the understanding.

• The difference between a pipeline and a true autonomous agent (it's the loop + memory + tool orchestration)
• How JSON tool schemas form the contract between an LLM and Python functions
• How sliding window memory keeps context within limits across multi-turn conversations
• How the sub-agent pattern works — an agent calling another agent internally
• How to migrate a terminal-based Python app into a production-ready web UI with Chainlit

---

📄 License

MIT License — feel free to use, fork, and build on this.

---

🙋 Author

Jainil — Computer Engineering Student
LinkedIn • GitHub

⭐ If this helped you understand how AI agents actually work under the hood, drop a star!