A research paper proposes a fully local AI cascade framework for de-identifying educational dialogue transcripts, which is sensitive to personally identifiable information (PII) and curricular content. The framework combines lightweight encoders with deterministic rules and a context-aware reviewer to make Redact/Keep decisions. The strongest local configuration achieves 0.958 macro F1, outperforming commercial API and LLM-only baselines, and runs entirely on a single laptop.