PII detection and redaction in Prompt Engineering / GenAI - Model Pipeline Trace

Model Pipeline - PII detection and redaction

This pipeline detects personal information in text and replaces it with safe placeholders. It helps protect privacy by automatically finding and hiding sensitive data like names, emails, and phone numbers.

Data Flow - 5 Stages

1Input Text

1000 sentences x variable length→Raw text containing personal information→1000 sentences x variable length

"Hello, my name is Alice and my email is [e-mail]."

↓

2Text Preprocessing

1000 sentences x variable length→Lowercase, remove extra spaces, tokenize sentences→1000 sentences x tokens per sentence

["hello", ",", "my", "name", "is", "alice", "and", "my", "email", "is", "[e-mail]", "."]

↓

3Feature Extraction

1000 sentences x tokens→Convert tokens to word embeddings (numeric vectors)→1000 sentences x tokens x 300 features

[[0.12, -0.05, ..., 0.33], ..., [0.45, 0.01, ..., -0.22]]

↓

4Model Prediction

1000 sentences x tokens x 300 features→Named Entity Recognition model tags tokens as PII or not→1000 sentences x tokens with PII tags

[('hello', 'O'), ('my', 'O'), ('name', 'O'), ('is', 'O'), ('alice', 'B-PER'), ('and', 'O'), ('my', 'O'), ('email', 'O'), ('is', 'O'), ('[e-mail]', 'B-EMAIL'), ('.', 'O')]

↓

5Redaction

1000 sentences x tokens with PII tags→Replace PII tokens with placeholders→1000 sentences x tokens with redacted text

"Hello, my name is [PERSON] and my email is [EMAIL]."

Training Trace - Epoch by Epoch

Loss
0.9 |***************
0.7 |************
0.5 |********
0.3 |*****
0.1 |**
    +----------------
     1 2 3 4 5 Epochs

Epoch	Loss ↓	Accuracy ↑	Observation
1	0.85	0.65	Model starts learning to identify PII with moderate accuracy.
2	0.60	0.78	Loss decreases and accuracy improves as model learns patterns.
3	0.45	0.85	Model shows good ability to detect PII entities.
4	0.35	0.90	Further improvement with more precise tagging.
5	0.30	0.92	Model converges with high accuracy and low loss.

Prediction Trace - 5 Layers

Layer 1: Input Text

Layer 2: Text Preprocessing

Layer 3: Feature Extraction

Layer 4: Model Prediction

Layer 5: Redaction

Model Quiz - 3 Questions

What happens to the text after the redaction stage?

A Text is translated to another language

B PII is replaced with placeholders

C Text is summarized

D Text is converted to audio

Key Insight

This visualization shows how a model learns to find and hide personal information in text. The training improves the model's ability to tag sensitive data, enabling automatic redaction to protect privacy.