Yadav-Ranjan - AI & Data Governance Expert

OpenAI/Claude Governance Implementation

This interactive demo showcases how different governance policies affect AI outputs and content moderation.

Content Safety Prompt Engineering Policy Enforcement

Live Demo

Python

def analyze_with_different_policies(prompt, policy_levels):
    """Demo showing how different policy levels affect AI outputs"""
    results = {}
    
    for policy in policy_levels:
        messages = [
            {"role": "system", "content": f"You are an AI assistant with {policy} safety guidelines."},
            {"role": "user", "content": prompt}
        ]
        
        response = client.chat.completions.create(
            model="gpt-4",
            messages=messages,
            temperature=0.7,
            max_tokens=150
        )
        
        # Get response content
        result = response.choices[0].message.content
        
        # Get moderation scores
        moderation = classify_content(result)
        
        results[policy] = {
            "response": result,
            "moderation_scores": moderation
        }
    
    return results

Azure AI Content Safety Implementation

This implementation demonstrates how to detect and moderate harmful content across multiple categories with configurable policy thresholds.

Content Moderation Policy Controls Azure AI

Live Demo

Try Content Safety Analysis

Enter text to analyze:

Policy Strictness:

Strict (Low threshold)

Moderate

Permissive (High threshold)

PII Bias Detection & Mitigation Tool

This advanced tool analyzes text for personally identifiable information (PII) and tests for potential algorithmic bias in how different demographic groups are represented and protected. It implements state-of-the-art privacy-preserving techniques including K-anonymity and Differential Privacy.

Privacy Protection Bias Detection Advanced Anonymization Compliance

PII Bias Detection & Mitigation

Enter sample text with PII:

Demographic Analysis:

Gender

Age

Potential Ethnicity

PII Detection Results:

PII Type	Value	Confidence	Demographic

Bias Analysis:

Potential bias detected: Non-uniform PII detection rates across demographic groups.

PII Detection Rate by Gender:

Male: 85%

Female: 92%

PII Sensitivity by Gender:

Male: 75%

Female: 68%

Mitigation Options:

Select Mitigation Strategy:

This tool analyzes text for personally identifiable information (PII) and tests for potential algorithmic bias in how different demographic groups are represented and protected.

Enter sample text with PII:

Demographic Analysis:

Gender

Age

Potential Ethnicity

Python

import re
            import spacy
            from nameparser import HumanName
            from typing import List, Dict, Any, Tuple
            
            # Load NLP model
            nlp = spacy.load("en_core_web_md")
            
            # PII detection patterns
            PII_PATTERNS = {
                'SSN': r'\b\d{3}-\d{2}-\d{4}\b',
                'EMAIL': r'\b[A-Za-z0-9._%+-]+@[A-Za-z0-9.-]+\.[A-Z|a-z]{2,}\b',
                'PHONE': r'\b\(?\d{3}\)?[-.\s]?\d{3}[-.\s]?\d{4}\b',
                'ADDRESS': r'\b\d+\s+[A-Za-z0-9\s,]+(?:Avenue|Lane|Road|Boulevard|Drive|Street|Ave|Dr|Rd|Blvd|Ln|St)\.?\b',
                'CREDIT_CARD': r'\b(?:\d{4}[- ]?){3}\d{4}\b'
            }
            
            def detect_pii(text: str) -> List[Dict[str, Any]]:
                """Detect PII in text with demographic analysis."""
                pii_found = []
                doc = nlp(text)
                
                # Detect names and demographic info
                for ent in doc.ents:
                    if ent.label_ == "PERSON":
                        name = str(ent)
                        parsed_name = HumanName(name)
                        # Attempt to infer gender (simplified)
                        gender = infer_gender(parsed_name.first)
                        pii_found.append({
                            'type': 'NAME',
                            'value': name,
                            'confidence': 0.85,
                            'demographic': gender
                        })
                
                # Detect other PII types using regex
                for pii_type, pattern in PII_PATTERNS.items():
                    for match in re.finditer(pattern, text):
                        pii_found.append({
                            'type': pii_type,
                            'value': match.group(),
                            'confidence': 0.92,
                            'demographic': find_associated_demographic(text, match.span())
                        })
                
                return pii_found
            
            def analyze_bias(pii_results: List[Dict], demographics: List[str]) -> Dict:
                """Analyze potential bias in PII detection across demographics."""
                bias_metrics = {}
                
                for demo in demographics:
                    demo_instances = [p for p in pii_results if p['demographic'].get('type') == demo]
                    if demo_instances:
                        for subgroup in set(d['demographic']['value'] for d in demo_instances):
                            group_key = f"{demo}:{subgroup}"
                            group_pii = [p for p in demo_instances 
                                         if p['demographic']['value'] == subgroup]
                            
                            bias_metrics[group_key] = {
                                'detection_rate': len(group_pii) / len([p for p in pii_results 
                                                           if p['demographic'].get('value') == subgroup]),
                                'avg_confidence': sum(p['confidence'] for p in group_pii) / len(group_pii),
                                'pii_types': {t: len([p for p in group_pii if p['type'] == t]) 
                                            for t in set(p['type'] for p in group_pii)}
                            }
                
                return bias_metrics
            
            def mitigate_bias(text: str, pii_results: List[Dict], 
                             strategy: str = 'uniform') -> str:
                """Apply bias mitigation strategy to PII handling."""
                mitigated_text = text
                
                if strategy == 'uniform':
                    # Replace all detected PII with consistent redaction
                    for pii in pii_results:
                        mitigated_text = mitigated_text.replace(
                            pii['value'], f"[{pii['type']}]")
                
                elif strategy == 'contextual':
                    # Apply different redaction based on PII type and context
                    # preserving demographic balance
                    for pii in pii_results:
                        replacement = get_contextual_replacement(pii)
                        mitigated_text = mitigated_text.replace(pii['value'], replacement)
                
                elif strategy == 'statistical':
                    # Ensure statistically equivalent privacy preservation
                    # across demographic groups
                    demographics = {}
                    for pii in pii_results:
                        demo = f"{pii['demographic'].get('type')}:{pii['demographic'].get('value')}"
                        if demo not in demographics:
                            demographics[demo] = []
                        demographics[demo].append(pii)
                    
                    # Balance protection across groups
                    for demo, items in demographics.items():
                        protection_level = calculate_protection_level(items)
                        for pii in items:
                            replacement = apply_protection_level(pii, protection_level)
                            mitigated_text = mitigated_text.replace(pii['value'], replacement)
                
                return mitigated_text

PII Bias Detection & Mitigation

This tool helps identify potential biases in how personally identifiable information (PII) is detected and processed across different demographic groups.

Key Features

Comprehensive PII Detection - Identifies names, addresses, SSNs, phone numbers, email addresses, and other sensitive information
Demographic Analysis - Examines detection patterns across gender, age, and other demographic factors
Bias Metrics - Quantifies differences in detection rates, confidence scores, and sensitivity levels
Mitigation Strategies - Implements techniques to ensure uniform privacy protection

Mitigation Strategies

Strategy	Description	Best For
Uniform PII Processing	Applies identical redaction to all identified PII regardless of demographic factors	Most situations where consistent handling is required
Contextual Privacy	Adjusts PII handling based on context while ensuring equivalent protection levels	When context-specific handling is needed without introducing bias
Statistical Parity	Ensures statistically equivalent privacy preservation across all demographic groups	Highly sensitive applications requiring formal fairness guarantees
Differential Privacy	Adds calibrated noise to numerical data to protect individuals while preserving statistical properties	Releasing aggregate statistics, machine learning, and data analytics with formal privacy guarantees

Note: This tool is intended for educational and demonstration purposes. In production environments, more comprehensive privacy and fairness assessments should be conducted.

AI & Data Governance Expert

AI Governance Expertise

Responsible AI

Bias Detection

Model Transparency

OpenAI/Claude Governance Implementation

Azure AI Content Safety Implementation

Try Content Safety Analysis

Analysis Results:

PII Bias Detection & Mitigation Tool

PII Bias Detection & Mitigation

PII Detection Results:

Bias Analysis:

Mitigation Options:

Mitigated Text:

Data Governance Expertise

Data Privacy

Data Quality

Data Catalogs

Data Lineage

Compliance Management