🧮 Day 22 — AI with Calculator Tool (Function/Tool Calling)

Part of my 30-Day AI Engineering Bootcamp Day 22 — building one AI project a day to become job-ready.

🤔 Why I Built This

LLMs are bad at exact math — they predict likely-looking text, they don't actually compute. Ask one 847 × 392 and it might confidently hallucinate a wrong number. Today I gave the AI a real calculator function it can call whenever it needs precise math, instead of guessing. This is also my first step into tool/function calling — the foundation every agent (Day 24-25) and tool-using AI system is built on.

🧠 Thought Process

1. Write a real Python function (calculate) that safely evaluates math expressions
2. Describe that function to the model in a structured JSON schema — name, description, parameters
3. Send the user's question + the tool description to Groq
4. Let the model decide: answer directly, OR ask to call the calculator
5. If it asks to call a tool, my code actually runs the function and gets the real result
6. Feed that real result back to the model so it can write the final natural-language answer

⚙️ What It Does

• 🧮 Calls a real calculator function for any math in the question
• 🧠 Answers general knowledge questions directly, without unnecessary tool calls
• 🔀 Can mix both in a single question (e.g., "what's 23×47, and what's the capital of France?")
• 🛡️ Safely sandboxes the calculator — only digits and math operators are allowed through, blocking any attempt to run arbitrary code

🛠️ Tech Stack

• Python — core logic
• Groq API (LLaMA 3.3 70B) — tool-calling enabled model
• Streamlit — simple UI

🚀 How to Run

git clone https://github.com/PrashikSawant/AI-with-Calculator-Tool
cd Day22-CalculatorTool
python -m venv venv
.\venv\Scripts\activate
pip install -r requirements.txt
# Add GROQ_API_KEY to .env
streamlit run app.py

💡 What I Learned

• LLMs don't calculate — they predict text. Real precision needs real code, not a better prompt
• Tool calling is a 4-step loop: describe tools → model decides → your code executes → model writes the final answer using the real result
• The model itself decides when a tool is needed — I never hardcoded "if it contains numbers, do math." It understood that from the tool's description alone