Fraud Model Explainability Assistant - Strands Agents

app.py

@@ -0,0 +1,837 @@
+"""
+Fraud Model Explainability Assistant - Strands Agents
+
+An AI-powered assistant that helps fraud analysts and executives understand
+why specific applications were flagged as fraudulent, translating complex
+model outputs into actionable insights.
+
+Use Cases:
+- Executive briefings on fraud decisions
+- Fair lending compliance documentation
+- Analyst investigation support
+- Model decision audit trails
+
+Author: Fraud Model Data Science Team
+"""
+
+import os
+import random
+from datetime import datetime, timedelta
+from typing import Optional
+import gradio as gr
+from strands import Agent, tool
+from strands.models.openai import OpenAIModel
+
+
+# =============================================================================
+# MOCK DATA GENERATORS
+# =============================================================================
+# In production, these would connect to your actual data systems
+# (e.g., Snowflake, feature store, model serving infrastructure)
+
+def generate_mock_application(app_id: str) -> dict:
+    """Generate realistic mock application data for demo purposes."""
+    random.seed(hash(app_id) % 2**32)
+    
+    risk_level = random.choice(["low", "medium", "high", "very_high"])
+    
+    base_data = {
+        "application_id": app_id,
+        "timestamp": (datetime.now() - timedelta(days=random.randint(0, 30))).isoformat(),
+        "portfolio": random.choice(["Retail Card", "Payment Solutions", "CareCredit"]),
+        "requested_credit_line": random.randint(500, 25000),
+        "fraud_score": {
+            "low": random.randint(150, 350),
+            "medium": random.randint(400, 550),
+            "high": random.randint(600, 750),
+            "very_high": random.randint(800, 950)
+        }[risk_level],
+        "fraud_score_percentile": {
+            "low": random.randint(5, 30),
+            "medium": random.randint(40, 60),
+            "high": random.randint(75, 90),
+            "very_high": random.randint(92, 99)
+        }[risk_level],
+        "decision": "FLAGGED" if risk_level in ["high", "very_high"] else "APPROVED",
+        "risk_level": risk_level,
+    }
+    
+    # Features that contribute to fraud scoring
+    if risk_level in ["high", "very_high"]:
+        base_data["features"] = {
+            "ssn_issue_date_vs_credit_age_mismatch": random.uniform(0.7, 0.95),
+            "device_velocity_30d": random.randint(5, 15),
+            "address_type": random.choice(["CMRA", "PO_BOX", "VACANT"]),
+            "phone_type": random.choice(["VOIP", "PREPAID"]),
+            "email_domain_age_days": random.randint(1, 30),
+            "application_velocity_14d": random.randint(3, 8),
+            "identity_linkage_count": random.randint(4, 12),
+            "credit_inquiry_spike": True,
+            "synthetic_id_score": random.uniform(0.75, 0.98),
+        }
+    else:
+        base_data["features"] = {
+            "ssn_issue_date_vs_credit_age_mismatch": random.uniform(0.0, 0.2),
+            "device_velocity_30d": random.randint(1, 2),
+            "address_type": "RESIDENTIAL",
+            "phone_type": "POSTPAID",
+            "email_domain_age_days": random.randint(365, 3650),
+            "application_velocity_14d": random.randint(0, 1),
+            "identity_linkage_count": random.randint(0, 2),
+            "credit_inquiry_spike": False,
+            "synthetic_id_score": random.uniform(0.05, 0.25),
+        }
+    
+    return base_data
+
+
+# =============================================================================
+# FRAUD EXPLAINABILITY TOOLS
+# =============================================================================
+
+@tool
+def get_application_summary(application_id: str) -> str:
+    """
+    Retrieve basic information about a credit application including
+    fraud score, decision, portfolio, and timestamp.
+    
+    Args:
+        application_id: The unique identifier for the application (e.g., "APP-12345")
+    
+    Returns:
+        A summary of the application details and fraud assessment
+    """
+    app = generate_mock_application(application_id)
+    
+    return f"""
+APPLICATION SUMMARY
+==================
+Application ID: {app['application_id']}
+Submission Date: {app['timestamp'][:10]}
+Portfolio: {app['portfolio']}
+Requested Credit Line: ${app['requested_credit_line']:,}
+
+FRAUD ASSESSMENT
+----------------
+Fraud Score: {app['fraud_score']} / 1000
+Risk Percentile: {app['fraud_score_percentile']}th percentile
+Risk Level: {app['risk_level'].upper()}
+Decision: {app['decision']}
+"""
+
+
+@tool
+def explain_fraud_score(application_id: str) -> str:
+    """
+    Get detailed SHAP-style feature attribution explanation for why an 
+    application received its fraud score. Shows which factors contributed
+    most to the risk assessment.
+    
+    Args:
+        application_id: The unique identifier for the application
+    
+    Returns:
+        Detailed breakdown of contributing factors with impact scores
+    """
+    app = generate_mock_application(application_id)
+    features = app["features"]
+    
+    # Simulate SHAP values (in production, these come from your model)
+    explanations = []
+    
+    if features["ssn_issue_date_vs_credit_age_mismatch"] > 0.5:
+        explanations.append({
+            "feature": "SSN Issue Date vs Credit Age Mismatch",
+            "value": f"{features['ssn_issue_date_vs_credit_age_mismatch']:.0%}",
+            "impact": "+187 points",
+            "direction": "INCREASES RISK",
+            "explanation": "SSN was issued recently but credit file shows longer history, a key synthetic ID indicator"
+        })
+    
+    if features["device_velocity_30d"] > 3:
+        explanations.append({
+            "feature": "Device Velocity (30 days)",
+            "value": f"{features['device_velocity_30d']} applications",
+            "impact": "+142 points",
+            "direction": "INCREASES RISK",
+            "explanation": "Same device fingerprint linked to multiple applications in short period"
+        })
+    
+    if features["address_type"] in ["CMRA", "PO_BOX", "VACANT"]:
+        explanations.append({
+            "feature": "Address Type",
+            "value": features["address_type"],
+            "impact": "+98 points",
+            "direction": "INCREASES RISK",
+            "explanation": f"Address classified as {features['address_type']} (Commercial Mail Receiving Agency or high-risk type)"
+        })
+    
+    if features["synthetic_id_score"] > 0.6:
+        explanations.append({
+            "feature": "Synthetic Identity Score",
+            "value": f"{features['synthetic_id_score']:.0%}",
+            "impact": "+156 points",
+            "direction": "INCREASES RISK",
+            "explanation": "Composite score from ensemble model indicates high probability of synthetic identity"
+        })
+    
+    if features["application_velocity_14d"] > 2:
+        explanations.append({
+            "feature": "Application Velocity (14 days)",
+            "value": f"{features['application_velocity_14d']} applications",
+            "impact": "+78 points",
+            "direction": "INCREASES RISK",
+            "explanation": "Multiple credit applications submitted in short timeframe"
+        })
+    
+    if features["email_domain_age_days"] < 60:
+        explanations.append({
+            "feature": "Email Domain Age",
+            "value": f"{features['email_domain_age_days']} days",
+            "impact": "+45 points",
+            "direction": "INCREASES RISK",
+            "explanation": "Email address created very recently"
+        })
+    
+    if features["phone_type"] in ["VOIP", "PREPAID"]:
+        explanations.append({
+            "feature": "Phone Type",
+            "value": features["phone_type"],
+            "impact": "+62 points",
+            "direction": "INCREASES RISK",
+            "explanation": "Non-traditional phone type associated with higher fraud rates"
+        })
+    
+    # If low risk, show protective factors
+    if app["risk_level"] == "low":
+        explanations = [
+            {
+                "feature": "Established Credit History",
+                "value": "12+ years",
+                "impact": "-120 points",
+                "direction": "DECREASES RISK",
+                "explanation": "Long credit history consistent with SSN issue date"
+            },
+            {
+                "feature": "Stable Contact Information",
+                "value": "Verified",
+                "impact": "-85 points",
+                "direction": "DECREASES RISK",
+                "explanation": "Phone and address verified with multiple data sources"
+            },
+            {
+                "feature": "Low Application Velocity",
+                "value": "1 in 90 days",
+                "impact": "-45 points",
+                "direction": "DECREASES RISK",
+                "explanation": "Normal application pattern"
+            }
+        ]
+    
+    # Format output
+    output = f"""
+FRAUD SCORE EXPLANATION
+=======================
+Application ID: {application_id}
+Final Fraud Score: {app['fraud_score']} / 1000
+Model: XGBoost Fraud Ensemble v3.2
+
+TOP CONTRIBUTING FACTORS (ranked by impact):
+--------------------------------------------
+"""
+    
+    for i, exp in enumerate(sorted(explanations, key=lambda x: abs(int(x["impact"].split()[0])), reverse=True), 1):
+        output += f"""
+{i}. {exp['feature']}
+   Value: {exp['value']}
+   Impact: {exp['impact']} ({exp['direction']})
+   → {exp['explanation']}
+"""
+    
+    return output
+
+
+@tool
+def compare_to_population(application_id: str, comparison_group: str = "approved") -> str:
+    """
+    Compare an application's features to the approved or denied population
+    to show how unusual the applicant's characteristics are.
+    
+    Args:
+        application_id: The unique identifier for the application
+        comparison_group: Either "approved" or "denied" population to compare against
+    
+    Returns:
+        Statistical comparison showing how the application differs from typical cases
+    """
+    app = generate_mock_application(application_id)
+    features = app["features"]
+    
+    # Mock population statistics
+    population_stats = {
+        "approved": {
+            "ssn_credit_mismatch_mean": 0.08,
+            "ssn_credit_mismatch_std": 0.12,
+            "device_velocity_mean": 1.2,
+            "device_velocity_std": 0.8,
+            "synthetic_score_mean": 0.15,
+            "synthetic_score_std": 0.10,
+            "app_velocity_mean": 0.5,
+            "app_velocity_std": 0.7,
+        },
+        "denied": {
+            "ssn_credit_mismatch_mean": 0.72,
+            "ssn_credit_mismatch_std": 0.18,
+            "device_velocity_mean": 6.5,
+            "device_velocity_std": 3.2,
+            "synthetic_score_mean": 0.78,
+            "synthetic_score_std": 0.15,
+            "app_velocity_mean": 4.2,
+            "app_velocity_std": 2.1,
+        }
+    }
+    
+    stats = population_stats.get(comparison_group, population_stats["approved"])
+    
+    def calc_z_score(value, mean, std):
+        if std == 0:
+            return 0
+        return (value - mean) / std
+    
+    comparisons = [
+        {
+            "feature": "SSN/Credit Age Mismatch",
+            "applicant_value": f"{features['ssn_issue_date_vs_credit_age_mismatch']:.0%}",
+            "population_mean": f"{stats['ssn_credit_mismatch_mean']:.0%}",
+            "z_score": calc_z_score(features['ssn_issue_date_vs_credit_age_mismatch'], 
+                                    stats['ssn_credit_mismatch_mean'], 
+                                    stats['ssn_credit_mismatch_std'])
+        },
+        {
+            "feature": "Device Velocity (30d)",
+            "applicant_value": str(features['device_velocity_30d']),
+            "population_mean": f"{stats['device_velocity_mean']:.1f}",
+            "z_score": calc_z_score(features['device_velocity_30d'], 
+                                    stats['device_velocity_mean'], 
+                                    stats['device_velocity_std'])
+        },
+        {
+            "feature": "Synthetic ID Score",
+            "applicant_value": f"{features['synthetic_id_score']:.0%}",
+            "population_mean": f"{stats['synthetic_score_mean']:.0%}",
+            "z_score": calc_z_score(features['synthetic_id_score'], 
+                                    stats['synthetic_score_mean'], 
+                                    stats['synthetic_score_std'])
+        },
+        {
+            "feature": "Application Velocity (14d)",
+            "applicant_value": str(features['application_velocity_14d']),
+            "population_mean": f"{stats['app_velocity_mean']:.1f}",
+            "z_score": calc_z_score(features['application_velocity_14d'], 
+                                    stats['app_velocity_mean'], 
+                                    stats['app_velocity_std'])
+        },
+    ]
+    
+    output = f"""
+POPULATION COMPARISON ANALYSIS
+==============================
+Application ID: {application_id}
+Comparison Group: {comparison_group.upper()} applications (last 12 months)
+Sample Size: {'847,293' if comparison_group == 'approved' else '23,847'} applications
+
+FEATURE COMPARISON:
+-------------------
+{"Feature":<30} {"Applicant":<15} {"Population Mean":<18} {"Z-Score":<10} {"Assessment"}
+{"-"*95}
+"""
+    
+    for comp in comparisons:
+        z = comp["z_score"]
+        if abs(z) > 3:
+            assessment = "⚠️  EXTREME OUTLIER"
+        elif abs(z) > 2:
+            assessment = "🔶 SIGNIFICANT DEVIATION"
+        elif abs(z) > 1:
+            assessment = "🔷 MILD DEVIATION"
+        else:
+            assessment = "✅ WITHIN NORMAL"
+        
+        output += f"{comp['feature']:<30} {comp['applicant_value']:<15} {comp['population_mean']:<18} {z:>+.2f}σ      {assessment}\n"
+    
+    # Summary
+    extreme_count = sum(1 for c in comparisons if abs(c["z_score"]) > 2)
+    
+    output += f"""
+SUMMARY:
+--------
+{extreme_count} of {len(comparisons)} features show significant deviation (|z| > 2σ) from {comparison_group} population.
+"""
+    
+    if extreme_count >= 2:
+        output += f"This application's profile is statistically unusual compared to typically {comparison_group} applications."
+    
+    return output
+
+
+@tool
+def check_fair_lending_flags(application_id: str) -> str:
+    """
+    Check for potential fair lending concerns in the fraud decision.
+    Reviews whether protected class proxies may have influenced the score
+    and provides compliance documentation.
+    
+    Args:
+        application_id: The unique identifier for the application
+    
+    Returns:
+        Fair lending compliance assessment and documentation
+    """
+    app = generate_mock_application(application_id)
+    
+    # Mock fair lending analysis
+    output = f"""
+FAIR LENDING COMPLIANCE REVIEW
+==============================
+Application ID: {application_id}
+Review Date: {datetime.now().strftime('%Y-%m-%d %H:%M:%S')}
+Model: XGBoost Fraud Ensemble v3.2
+
+PROTECTED CLASS PROXY ANALYSIS:
+-------------------------------
+The following features were analyzed for potential correlation with protected characteristics:
+
+✅ Geography-Based Features:
+   - ZIP code used only for velocity calculations, not scoring
+   - No direct geographic risk scoring applied
+   - Compliant with ECOA geographic restrictions
+
+✅ Name-Based Features:
+   - No name-based features used in fraud model
+   - Identity verification uses SSN/DOB only
+   
+✅ Age-Related Features:
+   - Credit age features measure account history, not applicant age
+   - SSN issuance analysis targets synthetic ID patterns, not age discrimination
+   - Model tested for age disparate impact: PASSED (adverse impact ratio: 0.94)
+
+⚠️  REVIEW ITEMS:
+-----------------
+"""
+    
+    if app["features"].get("phone_type") in ["VOIP", "PREPAID"]:
+        output += """
+• Phone Type Feature:
+  - VOIP/Prepaid flagged as risk factor
+  - Documented business justification: 73% of confirmed synthetic fraud uses VOIP
+  - Disparate impact testing: PASSED (ratio: 0.89)
+  - Alternative considered: None available with equivalent predictive power
+"""
+    
+    if app["features"].get("address_type") in ["CMRA", "PO_BOX"]:
+        output += """
+• Address Type Feature:
+  - CMRA/PO Box flagged as risk factor  
+  - Documented business justification: Required for synthetic ID detection
+  - Disparate impact testing: PASSED (ratio: 0.91)
+  - Accommodations: Manual review pathway available for legitimate CMRA users
+"""
+    
+    output += f"""
+MODEL VALIDATION STATUS:
+------------------------
+Last Disparate Impact Test: 2024-11-15
+Last Adverse Action Review: 2024-12-01
+Model Risk Rating: LOW
+SR 11-7 Compliance: COMPLIANT
+
+ADVERSE ACTION REASON CODES:
+----------------------------
+If this application is denied, the following reason codes apply:
+"""
+    
+    if app["decision"] == "FLAGGED":
+        reasons = [
+            "FA01 - Unable to verify identity information",
+            "FA03 - Inconsistent application information",
+            "FA07 - High-risk contact information patterns",
+        ]
+        for i, reason in enumerate(reasons, 1):
+            output += f"  {i}. {reason}\n"
+    else:
+        output += "  N/A - Application approved\n"
+    
+    output += """
+DOCUMENTATION:
+--------------
+This analysis is auto-generated for compliance documentation.
+Full model documentation available in Model Risk Management system.
+Contact: model-governance@company.com
+"""
+    
+    return output
+
+
+@tool
+def get_identity_network(application_id: str) -> str:
+    """
+    Analyze the identity linkage network for an application, showing
+    connections to other applications via shared attributes (device,
+    phone, email, address, SSN patterns).
+    
+    Args:
+        application_id: The unique identifier for the application
+    
+    Returns:
+        Network analysis showing linked applications and risk patterns
+    """
+    app = generate_mock_application(application_id)
+    features = app["features"]
+    
+    linkage_count = features.get("identity_linkage_count", 0)
+    
+    output = f"""
+IDENTITY NETWORK ANALYSIS
+=========================
+Application ID: {application_id}
+Analysis Date: {datetime.now().strftime('%Y-%m-%d')}
+
+LINKAGE SUMMARY:
+----------------
+Total Linked Applications: {linkage_count}
+"""
+    
+    if linkage_count > 3:
+        # Generate mock linked applications for high-risk cases
+        random.seed(hash(application_id) % 2**32)
+        
+        link_types = {
+            "device_fingerprint": random.randint(2, min(linkage_count, 8)),
+            "phone_number": random.randint(1, min(linkage_count, 4)),
+            "email_pattern": random.randint(1, min(linkage_count, 3)),
+            "address": random.randint(1, min(linkage_count, 5)),
+        }
+        
+        output += f"""
+LINKAGE BREAKDOWN:
+------------------
+• Device Fingerprint Links: {link_types['device_fingerprint']} applications
+• Phone Number Links: {link_types['phone_number']} applications  
+• Email Pattern Links: {link_types['email_pattern']} applications
+• Address Links: {link_types['address']} applications
+
+LINKED APPLICATION DETAILS:
+---------------------------
+"""
+        
+        statuses = ["CONFIRMED_FRAUD", "FLAGGED", "DENIED", "CHARGED_OFF", "APPROVED"]
+        weights = [0.3, 0.25, 0.2, 0.15, 0.1] if app["risk_level"] in ["high", "very_high"] else [0.05, 0.1, 0.15, 0.1, 0.6]
+        
+        for i in range(min(linkage_count, 6)):
+            linked_id = f"APP-{random.randint(10000, 99999)}"
+            link_type = random.choice(list(link_types.keys()))
+            status = random.choices(statuses, weights=weights)[0]
+            days_ago = random.randint(1, 180)
+            
+            status_emoji = {
+                "CONFIRMED_FRAUD": "🔴",
+                "FLAGGED": "🟠", 
+                "DENIED": "🟡",
+                "CHARGED_OFF": "🔴",
+                "APPROVED": "🟢"
+            }
+            
+            output += f"  {status_emoji.get(status, '⚪')} {linked_id} | {link_type.replace('_', ' ').title()} | {status} | {days_ago}d ago\n"
+        
+        # Risk assessment
+        fraud_links = sum(1 for _ in range(linkage_count) if random.random() < 0.4)
+        
+        output += f"""
+NETWORK RISK ASSESSMENT:
+------------------------
+• Confirmed Fraud in Network: {fraud_links} application(s)
+• Network Risk Score: {min(100, linkage_count * 12 + fraud_links * 25)}/100
+• Ring Pattern Detected: {"YES ⚠️" if linkage_count > 5 else "NO"}
+• Velocity Anomaly: {"YES ⚠️" if features.get('device_velocity_30d', 0) > 5 else "NO"}
+
+RECOMMENDATION:
+---------------
+{"⚠️  HIGH-RISK NETWORK - Manual review recommended" if linkage_count > 5 else "🔶 ELEVATED RISK - Monitor for additional activity"}
+"""
+    
+    else:
+        output += """
+LINKAGE BREAKDOWN:
+------------------
+• Device Fingerprint Links: 0-1 applications
+• Phone Number Links: 0 applications
+• Email Pattern Links: 0 applications
+• Address Links: 1 application (same household likely)
+
+NETWORK RISK ASSESSMENT:
+------------------------
+• Network Risk Score: LOW
+• No suspicious patterns detected
+• Normal application profile
+
+✅ No concerning identity network patterns identified.
+"""
+    
+    return output
+
+
+@tool
+def get_model_performance(model_name: str = "xgboost_fraud_v3.2", portfolio: str = "all") -> str:
+    """
+    Retrieve current performance metrics for a fraud detection model,
+    including precision, recall, KS statistic, and financial impact.
+    
+    Args:
+        model_name: Name of the fraud model (default: xgboost_fraud_v3.2)
+        portfolio: Portfolio to filter by ("Retail Card", "Payment Solutions", "CareCredit", or "all")
+    
+    Returns:
+        Model performance metrics and trends
+    """
+    output = f"""
+MODEL PERFORMANCE DASHBOARD
+===========================
+Model: {model_name}
+Portfolio: {portfolio.upper()}
+Reporting Period: Last 30 Days
+Generated: {datetime.now().strftime('%Y-%m-%d %H:%M:%S')}
+
+DETECTION METRICS:
+------------------
+                        Current     Prior Month     Δ Change
+Fraud Detection Rate:   87.3%       84.1%          +3.2% ✅
+Precision (PPV):        34.2%       31.8%          +2.4% ✅
+False Positive Rate:    2.1%        2.4%           -0.3% ✅
+KS Statistic:           0.72        0.69           +0.03 ✅
+Gini Coefficient:       0.81        0.78           +0.03 ✅
+AUC-ROC:                0.91        0.89           +0.02 ✅
+
+FINANCIAL IMPACT:
+-----------------
+                        Current     Prior Month     Δ Change
+Fraud Losses Prevented: $4.2M       $3.8M          +$400K ✅
+False Positive Cost:    $890K       $920K          -$30K  ✅
+Net Benefit:            $3.31M      $2.88M         +$430K ✅
+ROI:                    372%        317%           +55%   ✅
+
+VOLUME METRICS:
+---------------
+Applications Scored:    1,247,832
+High-Risk Flags:        26,847 (2.15%)
+Manual Reviews:         8,421
+Confirmed Fraud:        9,182
+"""
+    
+    if portfolio != "all":
+        output += f"""
+PORTFOLIO BREAKDOWN ({portfolio}):
+{'='*40}
+Applications:           {random.randint(200000, 500000):,}
+Fraud Rate:             {random.uniform(0.5, 1.2):.2f}%
+Detection Rate:         {random.uniform(82, 92):.1f}%
+"""
+    
+    output += """
+MODEL HEALTH:
+-------------
+✅ Feature Drift (PSI): 0.08 (threshold: 0.25)
+✅ Score Distribution: Stable
+✅ Latency P99: 45ms (SLA: 100ms)
+⚠️  Challenger Model: +2.1% lift in shadow mode - review scheduled
+
+TREND ALERT:
+------------
+📈 Synthetic ID fraud attempts up 23% MoM - model adapting well
+📉 First-party fraud stable at historical levels
+"""
+    
+    return output
+
+
+# =============================================================================
+# SYSTEM PROMPT
+# =============================================================================
+
+SYSTEM_PROMPT = """
+You are a Fraud Model Explainability Assistant for a major financial services company.
+Your role is to help fraud analysts, data scientists, and executives understand 
+fraud model decisions and their implications.
+
+You have access to tools that can:
+1. Retrieve application summaries and fraud scores
+2. Explain why applications received specific fraud scores (SHAP-style explanations)
+3. Compare applications to approved/denied populations statistically
+4. Check for fair lending compliance concerns
+5. Analyze identity networks and linkages
+6. Show model performance metrics
+
+When answering questions:
+- Be precise and data-driven
+- Highlight the most important risk factors first
+- Explain technical concepts in business terms when speaking to executives
+- Always mention fair lending implications when relevant
+- Provide actionable insights, not just data
+
+For flagged applications, structure your response as:
+1. Quick summary (score, decision, risk level)
+2. Top contributing factors
+3. How unusual this is compared to the population
+4. Any compliance considerations
+5. Recommended next steps
+
+Remember: Your explanations may be used in regulatory examinations and audits,
+so be accurate and thorough.
+""".strip()
+
+
+# =============================================================================
+# AGENT SETUP
+# =============================================================================
+
+def create_agent():
+    """Create and configure the Strands fraud explainability agent."""
+    openai_api_key = os.environ.get('OPENAI_API_KEY')
+    
+    tools = [
+        get_application_summary,
+        explain_fraud_score,
+        compare_to_population,
+        check_fair_lending_flags,
+        get_identity_network,
+        get_model_performance,
+    ]
+    
+    if openai_api_key:
+        model = OpenAIModel(
+            client_args={"api_key": openai_api_key},
+            model_id="gpt-4o",
+            params={"temperature": 0.1, "max_tokens": 2048}
+        )
+        return Agent(model=model, system_prompt=SYSTEM_PROMPT, tools=tools)
+    else:
+        # Default to Bedrock
+        return Agent(system_prompt=SYSTEM_PROMPT, tools=tools)
+
+
+def query(question: str) -> str:
+    """Process a question using the fraud explainability agent."""
+    try:
+        agent = create_agent()
+        result = agent(question)
+        return str(result)
+    except Exception as e:
+        return f"Error processing question: {str(e)}"
+
+
+# =============================================================================
+# GRADIO INTERFACE
+# =============================================================================
+
+def process_question(question: str) -> str:
+    """Wrapper for Gradio interface."""
+    return query(question)
+
+
+# Custom CSS for professional appearance
+custom_css = """
+.gradio-container {
+    font-family: 'Segoe UI', Tahoma, Geneva, Verdana, sans-serif;
+}
+.output-text {
+    font-family: 'Consolas', 'Monaco', monospace;
+    font-size: 14px;
+    line-height: 1.5;
+}
+"""
+
+# Create the interface
+with gr.Blocks(css=custom_css, title="Fraud Model Explainability Assistant") as iface:
+    gr.Markdown("""
+    # 🔍 Fraud Model Explainability Assistant
+    
+    An AI-powered assistant that helps you understand fraud model decisions, 
+    investigate flagged applications, and ensure fair lending compliance.
+    
+    **Capabilities:**
+    - Explain why applications were flagged (SHAP-style feature attribution)
+    - Compare applications to approved/denied populations
+    - Analyze identity networks and linkages
+    - Check fair lending compliance
+    - Review model performance metrics
+    """)
+    
+    with gr.Row():
+        with gr.Column(scale=2):
+            question_input = gr.Textbox(
+                label="Ask a Question",
+                placeholder="e.g., Why was application APP-12345 flagged as high risk?",
+                lines=3
+            )
+            submit_btn = gr.Button("🔍 Analyze", variant="primary")
+        
+    with gr.Row():
+        output = gr.Textbox(
+            label="Analysis Results",
+            lines=25,
+            show_copy_button=True
+        )
+    
+    gr.Markdown("### 💡 Example Questions")
+    
+    examples = gr.Examples(
+        examples=[
+            ["Why was application APP-78432 flagged as high risk?"],
+            ["Explain the fraud score for APP-12345 and compare it to approved applications"],
+            ["Check fair lending compliance for application APP-55555"],
+            ["Show me the identity network analysis for APP-78432"],
+            ["What's the current model performance for the Retail Card portfolio?"],
+            ["I need to present APP-99999 to the CCO. Give me a complete risk summary with compliance review."],
+            ["Compare APP-12345 to the denied population and explain if this looks like synthetic ID fraud"],
+        ],
+        inputs=question_input
+    )
+    
+    submit_btn.click(fn=process_question, inputs=question_input, outputs=output)
+    question_input.submit(fn=process_question, inputs=question_input, outputs=output)
+    
+    gr.Markdown("""
+    ---
+    *Powered by Amazon Strands Agents SDK | Demo with synthetic data*
+    
+    **Note:** This demo uses mock data. In production, tools would connect to:
+    - Feature Store / Data Warehouse
+    - Model Serving Infrastructure  
+    - SHAP/Model Interpretation Services
+    - Compliance Documentation Systems
+    """)
+
+
+# =============================================================================
+# MAIN
+# =============================================================================
+
+if __name__ == "__main__":
+    import sys
+    
+    if "--demo" in sys.argv:
+        print("\n" + "="*70)
+        print("FRAUD MODEL EXPLAINABILITY ASSISTANT - DEMO")
+        print("="*70 + "\n")
+        
+        test_questions = [
+            "Why was application APP-78432 flagged as high risk?",
+            "What's the model performance for Retail Card?",
+        ]
+        
+        for q in test_questions:
+            print(f"Question: {q}")
+            print("-" * 50)
+            print(query(q))
+            print("\n")
+    else:
+        iface.launch()