Under HIPAA, GDPR, and SOC2 Type II, sending raw patient records, employee Social Security numbers, or customer credit card details to third-party cloud APIs can result in severe compliance fines.
The solution is an in-memory PII Sanitization Gateway: identifying and replacing sensitive entities with synthetic tokens (<PERSON_1>, <SSN_1>) on your local server before making the LLM request.
1. Reversible Masking Architecture
User Prompt (with PII)
│
▼
┌─────────────────────────────┐
│ Local Presidio Analyzer │ ➔ Identifies "John Doe", "4412-XXXX"
└──────────────┬──────────────┘
│
▼
┌─────────────────────────────┐
│ Local Presidio Anonymizer │ ➔ Replaces with <PERSON_1>, <CARD_1>
└──────────────┬──────────────┘
│ (Stores reverse mapping in RAM)
▼
┌─────────────────────────────┐
│ Cloud LLM (Gemini / OpenAI) │ ➔ Receives zero private customer PII!
└──────────────┬──────────────┘
│
▼
┌─────────────────────────────┐
│ Local Deanonymizer │ ➔ Replaces <PERSON_1> back with "John Doe"
└──────────────┬──────────────┘
│
▼
Clean Response Returned to User
2. Complete Python Implementation
from presidio_analyzer import AnalyzerEngine
from presidio_anonymizer import AnonymizerEngine
from presidio_anonymizer.entities import OperatorConfig
analyzer = AnalyzerEngine()
anonymizer = AnonymizerEngine()
def mask_customer_prompt(raw_text: str):
# 1. Detect PII entities
results = analyzer.analyze(
text=raw_text,
language="en",
entities=["PERSON", "EMAIL_ADDRESS", "PHONE_NUMBER", "US_SSN", "CREDIT_CARD"]
)
# 2. Anonymize with synthetic replace tokens
anonymized = anonymizer.anonymize(
text=raw_text,
analyzer_results=results,
operators={"DEFAULT": OperatorConfig("replace", {"new_value": "<REDACTED>"})}
)
return anonymized.text
# Test input
sample = "Patient John Doe (SSN: 000-12-3456) reported dizziness after taking medication."
clean = mask_customer_prompt(sample)
print(clean)
# Output: Patient <REDACTED> (SSN: <REDACTED>) reported dizziness after taking medication.
WEEKLY NEWSLETTER
Get Weekly AI Architect Cost & Strategy Updates
Join 14,000+ developers receiving weekly, data-driven cost-reduction blueprints and production-ready agent guidelines.
