PII Detection v1.1

Overview

Tork's PII Detection engine scans text through a 5-layer pipeline before it reaches AI models or is stored in logs. v1.1 introduces three major capabilities:

• Entity Slot Detection Engine — detects names by sentence structure (labels, titles, self-identification), not dictionary lookup. Works with African, Arabic, Indian, Chinese, and compound names that traditional approaches miss.
• Composable Regional PII Profiles — activate only the country and industry patterns you need via the optional region and industry parameters. Zero overhead when not used.
• Supersession logic — wider regional matches correctly override shorter L0 matches (e.g., Emirates ID supersedes a partial phone match).

Detection Pipeline

PII detection runs through 5 layers in sequence. Each layer only fires on text regions not already matched by an earlier layer, preventing double-counting. All positions are computed against the original text, and a single-pass redaction is applied at the end.

Regional PII Profiles

Activate country-specific and industry-specific PII patterns by passing the optional region and industry parameters. When omitted, only the 29 universal L0 patterns run — fully backward compatible with zero overhead.

Profiles are composable: pass multiple regions and an industry to build exactly the detection surface you need. For example, region: ["ae", "in"], industry: "healthcare" activates UAE, India, and healthcare patterns together.

Available Regions

Available Industries

Composability

Combine multiple regions and an industry in a single request. Each profile adds its patterns to the detection surface without duplicating universal L0 patterns.

# UAE + India + Healthcare: detects Emirates ID, Aadhaar, PAN,
# ICD-10 codes, MRN, and all 29 universal patterns
result = tork.govern(
    content="Emirates ID: 784-1234-1234567-1, PAN: ABCDE1234F, ICD-10: J45.0",
    region=["ae", "in"],
    industry="healthcare"
)

# Nigeria + Finance: detects NIN, BVN, CUSIP, ISIN,
# plus all universal patterns
result = tork.govern(
    content="NIN: 12345678901, CUSIP: 037833100",
    region=["ng"],
    industry="finance"
)

Cross-Cultural Name Detection

Traditional dictionary-based name detection fails for non-Western names. No dictionary can contain every Nigerian, Arabic, Indian, or Chinese name — and names from these cultures often have compound forms, honorifics, and structures that don't match Western first/last patterns.

Tork solves this with the Entity Slot Detection Engine (L1), which detects names by analyzing sentence structure rather than relying on dictionary lookup. It identifies the “slot” where a name appears and extracts whatever occupies that slot.

Before & After

This approach works with names from any culture without requiring the name to exist in a dictionary:

• African names: Chukwuemeka Okonkwo, Oluwaseun Adeyemi, Ngozi Okafor
• Arabic names: Abd al-Rahman al-Farsi, Sheikh Mohammed bin Rashid
• Indian names: Venkatanarasimharajuvaripeta, Subramaniam Chandrasekhar
• Chinese names: Zhang Wei, Li Xiaoming
• Compound names: Mary-Jane Watson, Jean-Pierre Dupont

Slot Pattern Categories

Code Examples

Integrate PII detection into your application with a single API call.

import tork

# Basic (universal patterns only)
result = tork.govern(content="My SSN is 123-45-6789")

# Regional profiles
result = tork.govern(
    content="Emirates ID: 784-1234-1234567-1, Phone: +971 50 123 4567",
    region=["ae"]
)

# Multi-region + industry
result = tork.govern(
    content="Aadhaar: 1234 5678 9012, ICD-10: J45.20",
    region=["in"],
    industry="healthcare"
)

# Access results
print(result["action"])        # "redact"
print(result["output"])        # redacted text
print(result["pii_detected"])  # [{"type": "aadhaar", "count": 1}, ...]

Performance

PII detection is optimized for production workloads with multiple performance layers.

Scanning for PII

Scan text to detect PII entities without redacting:

from tork_governance import TorkClient

client = TorkClient()

# Scan text for PII
result = client.pii.scan(
    text="Contact John Smith at john@example.com or 555-123-4567",
    entity_types=["EMAIL", "PHONE", "PERSON"]
)

for entity in result.entities:
    print(f"Found {entity.type}: {entity.text} (confidence: {entity.confidence})")
# Output:
# Found PERSON: John Smith (confidence: 0.95)
# Found EMAIL: john@example.com (confidence: 0.99)
# Found PHONE: 555-123-4567 (confidence: 0.98)

Redacting PII

Automatically redact detected PII with customizable replacement patterns:

# Redact PII from text
result = client.pii.redact(
    text="My SSN is 123-45-6789 and my email is jane@company.com",
    redaction_style="placeholder"  # or "mask", "hash", "synthetic"
)

print(result.redacted_text)
# Output: "My SSN is [SSN] and my email is [EMAIL]"

# Use masking style
result = client.pii.redact(
    text="Card: 4111-1111-1111-1111",
    redaction_style="mask"
)
print(result.redacted_text)
# Output: "Card: ****-****-****-1111"

Supported Entity Types

Country-Specific Patterns

Enable country-specific PII patterns for localized detection via the region parameter:

# Enable UK-specific patterns
result = tork.govern(
    content="My NI number is AB 12 34 56 C and sort code is 12-34-56 for my bank account",
    region=["gb"]
)

# Enable multiple regions
result = tork.govern(
    content="Emirates ID: 784-1234-1234567-1, Aadhaar: 1234 5678 9012",
    region=["ae", "in"]
)

# Supported regions: AU, US, GB, EU, AE, SA, NG, IN, JP, CN, KR, BR