Extremist Detection

Problem

Content moderation pipelines often struggle to separate harmful intent from neutral discussion. This project focused on building a classifier that prioritizes precision while preserving acceptable recall.

Approach

• Built a clean preprocessing layer for token normalization.
• Trained baseline and tuned models for comparison.
• Used error analysis to reduce false positives on ambiguous phrasing.

Results

The final model produced stable performance across validation folds and was wrapped in a reusable inference utility for downstream moderation workflows.

What I learned

• Label quality matters more than model complexity in early iterations.
• Threshold tuning is critical for practical moderation outcomes.
• Clear failure-case analysis speeds up model improvement.