Project
Math Mentor AI
An LLM and SymPy-based math-reasoning pipeline that independently verifies generated answers before presenting them to users.
Why I Built It
LLMs are fluent but not always correct — especially at math. I wanted to see if I could build a system that catches wrong answers before the user sees them, turning a language model into something more like a reliable calculator.
The Problem
LLM-generated math answers can look convincing while containing subtle reasoning errors.
The Solution
The pipeline separates answer generation from verification by using structured outputs and SymPy checks before presenting final explanations.
Results
The architecture provides a clearer path for catching incorrect generated answers and improving trust in math assistance.
What This Demonstrates
LLM workflow design, structured output parsing, symbolic verification integration, and separation of generation from verification for increased reliability.
Limitations
Symbolic verification coverage depends on problem type and how well the model expresses intermediate steps.
Future Improvements
Expand supported math domains and add confidence reporting for verification outcomes.