Certified Algorithmic Bias Auditor
CABA is designed for detecting, assessing, mitigating, documenting, and governing fairness risks in AI systems. It treats fairness as a socio-technical audit problem shaped by data, models, institutions, regulation, and downstream impact.
Algorithmic Bias
Understand where unfairness originates across data generation, modeling choices, and institutional decision logic.
- Data Bias
- Label Bias
- Sampling Bias
- Model Bias
Fairness Metrics
Measure bias through group-wise comparisons, error rates, selection rates, and fairness-performance trade-offs.
- Demographic Parity
- Equalized Odds
- Equal Opportunity
- Predictive Parity
Sensitive Attributes & Proxies
Detect indirect discrimination where apparently neutral variables behave as proxies for protected attributes.
- ZIP Code
- Education
- Language
- Device
- Proxy Discrimination
Data Auditing
Audit bias before training by testing representation, coverage, quality, labels, and dataset construction assumptions.
- Data Profiling
- Coverage Analysis
- Label Validation
Model Bias: Classical ML
Evaluate group-wise model errors across classical machine learning systems and structured prediction pipelines.
- Logistic Regression
- Decision Tree
- Random Forest
- XGBoost
Deep Learning Bias
Study latent and amplified bias in neural networks where representations may encode hidden social or statistical skew.
- Amplification
- Calibration Errors
- Latent Bias
NLP & LLM Bias
Audit language-level bias from corpora, embeddings, alignment procedures, RLHF skew, and generative outputs.
- Embedding Bias
- Corpus Bias
- RLHF Skew
Generative AI Risk
Assess synthetic harm created by generated content, stereotyping, misinformation, and unsafe model behaviour.
- Stereotyping
- Disinformation
- Synthetic Harm
Explainable AI
Use explanations to support human review, decision transparency, auditability, and contestability of model outcomes.
- LIME
- SHAP
- Counterfactuals
Bias Mitigation
Apply correction strategies at different stages of the pipeline while documenting residual risk and trade-offs.
- Pre-processing
- In-processing
- Post-processing
- Threshold Optimization
Algorithmic Impact Assessment
Perform pre-deployment risk checks before AI systems affect people, decisions, access, or regulated outcomes.
- Risk Scoring
- Approval Review
- Deployment Restriction
AI Governance
Connect models to external accountability through regulation, governance controls, documentation, and evidence systems.
- EU AI Act
- GDPR
- NIST AI RMF
- Regulatory Evidence
Audit Artifacts
Prepare the practical evidence auditors inspect: documentation, logs, traces, reports, and accountable decision records.
- Data Cards
- Model Cards
- Logs
- Decision Traces
Human-in-the-Loop
Design final accountability through review, escalation, domain judgment, and human control over consequential AI outputs.
- Human Review
- Escalation
- Final Accountability
Continuous Monitoring
Track fairness after deployment because bias can drift as live data, user behaviour, and institutional conditions change.
- Data Drift
- Error Drift
- Fairness Drift
- Re-audit