MATLAB to Python Agent with Verified Output

Problem

Teams moving off MATLAB often have hundreds of analysis scripts. Translating them with an LLM fails quietly: the Python runs but the answers are slightly wrong, because of 1-based indexing, column-major reshape, or different function defaults such as filtfilt edge padding. Code that looks right is not good enough; it needs numerical proof.

Approach

• Reference outputs come from a real MATLAB run and are committed, so MATLAB is not needed afterwards.
• The agent scans the script for known built-ins and adds only the relevant rows of a 50-function mapping table to the prompt.
• A generated pytest file compares each variable with np.allclose, with per-variable tolerances.
• On failure, the next prompt carries the previous code and the exact failing checks.

Results

• Five bundled samples were checked against MATLAB R2025b at tight default tolerances, using 6 API calls in total.
• One sample failed first because the model left MATLAB indexing in the code. The agent fed the error back, and the second attempt passed every check.
• Mock mode plants realistic bugs, such as row-major reshape, so the retry path is tested without API calls.

Limitations

• Numerical checks need one MATLAB run to produce reference values.
• It handles one script at a time; multi-file projects, classes and Simulink are out of scope.
• Five samples show the loop works, not that it would reach 100% on real-world code.
• Converted scripts run locally with no sandbox, so only convert code you trust.

Built with

• Python
• NumPy
• SciPy
• Matplotlib
• pytest
• MATLAB
• Streamlit
• Gemini
• Anthropic
• OpenAI