AI Text Detection System

The problem

Teachers, editors, and platforms want a quick screen for AI-generated text without sending content to a third-party API. The goal was an on-device, explainable classifier — not a black box.

The approach

• Feature engineering: type-token ratio, burstiness, sentence-length variance, punctuation density, function-word frequency.
• Trained Logistic Regression, Random Forest, and Gradient Boosting; compared with cross-validated ROC-AUC.
• Streamlit demo shows per-feature contributions so users can see why a passage was flagged.

Outcome

• Best model reached ~0.92 ROC-AUC on the held-out test set.
• Explainability was the headline feature — users trusted a transparent score over a single confidence number.