Redact or Keep? A Fully Local AI Cascade for Educational Dialogue De-Identification

In three linesLocal de-identification framework for educational dialogues. Two-stage cascade: union proposer (lightweight encoders + deterministic rules) generates PII candidates, then binary Redact/Keep reviewer uses dialogue context and speaker role. Achieves 0.958 macro F1 on math tutoring transcripts, outperforms commercial API (0.706) and local LLM baseline (0.767), runs on single laptop.

Summary generated by Claude — human-verified