Email Classification & PII Masking API

A FastAPI service that masks PII/PCI from support emails and classifies them into one of four categories: Incident, Request, Change, or Problem.

Architecture

Input Email → PII Masker (Regex + spaCy NER) → TF-IDF + LinearSVC Classifier → JSON Response

Project Structure

.
├── main.py                          # FastAPI app entry point & /classify router
├── models.py                        # ML model training and inference logic
├── utils.py                         # PII/PCI masking utilities
├── train_model.py                   # Script to train and save the model
├── email_classifier.joblib          # Serialised trained model (generated by train_model.py)
├── combined_emails_with_natural_pii.csv  # Training dataset
├── requirements.txt
└── README.md

PII Entities Detected

Entity	Method	Example
full_name	Regex trigger phrases + spaCy NER	"My name is John Doe"
email	Regex	user@example.com
phone_number	Regex	+1-555-123-4567
dob	Regex	15/06/1990
aadhar_num	Regex	1234 5678 9012
credit_debit_no	Regex	4111 1111 1111 1111
cvv_no	Regex	CVV: 123 → masks 123
expiry_no	Regex	12/25

Setup & Installation

python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt
python -m spacy download en_core_web_sm

Training the Model

python train_model.py
# Optionally specify paths:
python train_model.py --csv combined_emails_with_natural_pii.csv --model email_classifier.joblib

Running the API Locally

uvicorn main:app --host 0.0.0.0 --port 7860

API Usage

POST /classify

Request:

{
  "input_email_body": "Subject: Server Down\n\nMy name is John Smith. The server is unreachable. You can reach me at john@company.com."
}

Response:

{
  "input_email_body": "Subject: Server Down\n\nMy name is John Smith...",
  "list_of_masked_entities": [
    {
      "position": [33, 43],
      "classification": "full_name",
      "entity": "John Smith"
    },
    {
      "position": [89, 105],
      "classification": "email",
      "entity": "john@company.com"
    }
  ],
  "masked_email": "Subject: Server Down\n\nMy name is [full_name]. The server is unreachable. You can reach me at [email].",
  "category_of_the_email": "Incident"
}

GET /health

Returns {"status": "ok"}.

Model Performance

Trained on ~24,000 multilingual support emails. Test set (15%):

Class	Precision	Recall	F1
Change	0.94	0.89	0.92
Incident	0.73	0.86	0.79
Problem	0.62	0.43	0.51
Request	0.94	0.94	0.94
Accuracy			0.79

Hugging Face Spaces Deployment

Deploy as a Docker Space (no Gradio/Streamlit frontend):

POST https://<username>-<space-name>.hf.space/classify

The email_classifier.joblib model file must be committed to the Space repository.