PII Identifier

Context

Teams that handle customer data, forms or scanned documents often need to strip personal information before sharing files. Cloud redaction services add a dependency and a data-transfer step that is sometimes not acceptable.

PII Identifier runs entirely on the user's machine and handles three input types: CSV tables, free text, and images.

My contribution

• Built the detection and redaction web app using Microsoft Presidio and spaCy for text entities, custom policy rules, Tesseract OCR for text inside images, and OpenCV face detection.
• Authored a redaction policy covering 20+ categories of personal information.
• Scaled training and evaluation data with a seeded synthetic-corpus generator: 440 documents and 793 entity labels.
• Converted 100+ hand-adjudicated detections into documented keep/redact rules.

Process

1. 01

   Define the policy

   What counts as PII, what is kept, what is masked.

2. 02

   Text pipeline

   Microsoft Presidio and spaCy for text entities, extended with custom policy rules for IDs, fixed formats and context.

3. 03

   Tables and images

   Column-aware scanning for CSV files, Tesseract OCR for text inside images, and MediaPipe face detection with an OpenCV fallback.

4. 04

   Gold evaluation set

   Hand-labelled documents with held-out test sets, scored on what must be redacted and what must stay visible.

5. 05

   Synthetic corpus

   A seeded generator that produces labelled documents for training a custom entity model.

6. 06

   Adjudicate and document

   Detections reviewed by hand, and the decisions turned into documented keep/redact rules.