IBM AI Fairness (AIF360)

🏆 The Swiss Army knife for fairness analysis

What it is:
The most comprehensive open-source framework for bias detection and mitigation. Developed by IBM Research, scientifically validated, production-tested.

What we do with it:

• Calculate 70+ fairness metrics
  • Demographic Parity
  • Equal Opportunity
  • Equalized Odds
  • Calibration
  • Individual Fairness
• Implement bias mitigation
  • Pre-processing: Reweighing, Disparate Impact Remover
  • In-Processing: Prejudice Remover, Adversarial Debiasing
  • Post-processing: Calibrated Equalized Odds, Reject Option
• • Multi-attribute combinations
  • Identify hidden subgroups

For which use cases:

• HR recruiting (Annex III.4)
• Credit scoring (Annex III.5b)
• Any system with protected attributes

Our advantage:
We are one of the few German providers with in-depth AIF360 expertise.

Technical details:

• Python library
• Scikit-learn compatible
• TensorFlow/PyTorch integration
• 10+ datasets included (for benchmarking)

Output example:
→ Disparate Impact Ratio: 0.65 (critical, <0.80)
→ Equal Opportunity Difference: 0.15 (problematic)
→ Recommendation: Reweighing + Threshold Optimization
→ Expected Improvement: DI 0.65 → 0.82

Custom German NLP bias probes

🇩🇪 German language, German bias patterns

What it is:
Proprietary tools for bias detection in German language models.

What we do with it:

• Stereotype detection in BERT models
• "The doctor" vs. "The nurse"
• Gendered language bias
• Sentiment fairness for German texts
• Name-based bias (German vs. non-German names)

For which use cases:

• Chatbots (Customer Service)
• Content moderation
• Sentiment analysis
• Text Classification

Our advantage: The
German language has specific bias patterns (gendered nouns, Sie/Du). International tools overlook these.

Technical basis:

• German BERT (bert-base-german-cased)
• Custom Probe Tasks
• Stereotype datasets (self-curated)

Market differentiation:

Unique in Germany. No other providers have German NLP bias expertise at this level.

Microsoft Fairlearn

🔧 Constraint-based fairness optimization

What it is:
Practical fairness engineering framework with a focus on trade-off analyses.

What we do with it:

• Define fairness constraints
• "Demographic parity must be >90%"
• Equal Opportunity Difference <0.05
• Grid search for fair hyperparameters
• Visualize trade-off analyses
• Fairness vs. Accuracy Pareto Front
• Explicit cost-benefit analysis

When we use Fairlearn:

• For faster prototypes (simpler than AIF360)
• For client workshops (better visualizations)
• For scikit-learn pipelines (seamless integration)

Our advantage:
Combination with AIF360 for the best of both worlds.

Output example:
→ Baseline: 85% accuracy, 0.70 demographic parity
→ Option A: 83% accuracy, 0.85 DP (−2% accuracy, +15% fairness)
→ Option B: 80% accuracy, 0.95 DP (−5% accuracy, +25% fairness)
→ Your decision: Which trade-off do you accept?

SHAP (SHapley Additive exPlanations)

💡 The gold standard for model explainability

What it is: A
scientifically sound framework based on Shapley values from cooperative game theory. NIPS 2017 Best Paper Award.

What we do with it:

• Global Feature Importance
• Which features are most important?
• How strong is the influence?
• Local Explanations
• Why was this decision made?
• Feature-by-feature contribution
• Feature Interactions
• Visualizing nonlinear effects
• SHAP Interaction Values

For which use cases:

• Every ML model (tree-based, neural networks, linear)
• Particularly critical: HR, credit, healthcare

German market requirement:
“Why did the system make that decision?” is not an optional question in the German B2B market. It is an expectation. SHAP provides the answer.

EU AI Act Compliance:
Article 13: “High-risk AI systems shall be designed and developed… to enable users to interpret the system’s output”
→ SHAP technically meets this requirement.

Output examples:

• Global: "Income accounts for 35% of credit score"
• Local: "Your application was rejected due to: Income (−15 points), Age (−8 points), Credit History (−12 points)"
• Counterfactual: "With €5k more income, approval would be likely."

Technical details:

• Model-agnostic (works for almost everything)
• Fast approximations (kernel SHAP, tree SHAP)
• GPU acceleration possible
• Integration: Python, R, Spark

LIME (Local Interpretable Model-agnostic Explanations)

🔍 Backup & Complementary to SHAP

What it is:
Model-agnostic explanations through local linear approximation.

When we use LIME instead of SHAP:

• Text Explanations (Which words were crucial?)
• Image Explanations (Which pixel regions?)
• Highly complex ensembles (where SHAP is too slow)

Our approach:
SHAP = Primary, LIME = Validation & Special Cases

Output example:
Text classification: “This text was classified as ‘negative’ because of: ‘bad’ (0.45), ‘disappointing’ (0.32), ‘never again’ (0.28)”

Model Card Toolkit

📋 Standardized model documentation

What it is:
Framework for transparent, structured model documentation (Google/TensorFlow).

What we create with it:

• Intended Use & Limitations
• Training Data & Preprocessing
• Performance Metrics (total & per group)
• Fairness Metrics
• Considerations of ethics

EU AI Act relevance:
Model Cards directly comply with Article 13 transparency requirements.

Our service:
We create production-ready model cards for your systems.

Great Expectations

🗂️ Data Quality Engineering for AI

What it is:
The leading framework for data validation, profiling, and documentation.

What we do with it:

• Define data quality tests
• "Missing Values <5%"
• "Age between 18-100"
• Income distribution matches training
• Automated testing in pipelines
• Audit trails (when was what tested?)
• Data Docs (automatic documentation)

EU AI Act Article 10:
“Training, validation, and testing data sets shall be relevant, sufficiently representative, and, to the best extent possible, free of errors and complete.”

→ Great Expectations makes this requirement measurable.

Our advantage:

• Custom Expectation Suites for AI Act Compliance
• Integration with your ML pipelines
• Audit-ready documentation out of the box

Output examples:

• Data Quality Scorecard: 87/100 (good, but room for improvement)
• 3 Critical Issues: Missing Values in Protected Attributes
• 12 Warnings: Outliers in Income Feature
• Recommendation: Implement data cleaning pipeline

Technical details:

• Python-native
• SQL database support
• Spark-compatible (big data)
• Cloud-ready (AWS, GCP, Azure)

Alibi Detect

🚨 Production monitoring for ML systems

What it is:
State-of-the-art framework for drift, outlier, and adversarial detection. Developed by Seldon.

What we do with it:

• Data drift detection
• Kolmogorov-Smirnov test
• Maximum Mean Discrepancy
• Chi-squared test
• Concept Drift Detection
• Outlier detection
• Isolation Forest
• Variational autoencoders
• Adversarial detection
• Adversarial AE Detector

EU AI Act Article 72:
“Providers shall establish and document a post-market monitoring system”
→ Alibi Detect is the technical implementation.

Our service:

• Monitoring setup for your systems
• Custom drift detectors
• Integration with alerting (Slack, email, PagerDuty)

Output examples:

• Data drift score: 0.35 (moderate, attention required)
• Feature "Age" drifts significantly (p<0.001)
• Prediction distribution: 15% shift to the right
• Recommendation: Evaluate model retraining

Evidently AI

📊 Visualization & Reporting

What it is:
User-friendly dashboards and reports for ML monitoring.

Why in addition to Alibi Detect:

• Better visualizations (for non-technical stakeholders)
• Interactive HTML reports
• Pre-built dashboards

Our approach:
Alibi Detect = Detection Engine
Evidently = Visualization Layer

Output:
Attractive, shareable reports for management and auditors.

IBM ART (Adversarial Robustness Toolbox)

🛡️ Security Testing for AI

What it is:
Framework for adversarial attacks and defenses.

When we use it:

• Red teaming for critical systems
• Security audits
• Robustness testing

Use cases:

• Fraud Detection (Adversarial Environment)
• Autonomous systems (safety-critical)
• Face Recognition (Security-critical)

Service level:
ENTERPRISE only (Premium Service)

Output:

Success rate of adversarial attacks, defense recommendations

Captum (PyTorch Explainability)

🖼️ Computer Vision Explainability

What it is:
Explainability for PyTorch models, specializing in CV.

When we use it:

• Quality Inspection Systems (Industry 4.0)
• Medical imaging
• Autonomous Vehicles

Output:
Visualizations: “This pixel region led to the classification”