This competition tests your ability to engineer reliable agents using constrained, open-source models. To ensure a level playing field, the following technical constraints are strictly enforced:

1. Model Size Limit (< 10B Parameters)
You are not allowed to use massive frontier models like GPT-4o, Claude 3.5 Sonnet, or Llama 3 70B. You must use a model with fewer than 10 Billion parameters.

Recommended Models: llama-3.1-8b-instant or gemma2-9b-it.

API Gateway: You must access these models via the Groq API.

2. Mandatory Prompt Engineering Techniques
Because smaller models struggle with complex tool-calling loops, your code must explicitly demonstrate the use of:

System Prompts: You must define a strict persona and operational bounds for the billing agent.

Few-Shot Prompting: You must provide the model with at least one complete example of the "Thought -> Action -> Observation" loop in your prompt so it knows how to sequence the three math tools.

3. Tool Usage Enforcement
Your agent is strictly forbidden from doing the math "in its head." The final answer must be the result of the LLM passing the extracted variables through the three Python tools you created.

4. Code Review
This is a Code Competition. Winning submissions will be manually reviewed by the Lead Instructor after the deadline. If your notebook does not hit the Groq API, uses an oversized model, or bypasses the tool creation requirement, your submission will be disqualified.