Building Trust in AI Systems.

Oyekemi Abioye | Data Scientist & Responsible AI Advocate

Developing machine learning solutions with a strong emphasis on ethical considerations.

👋 About Me

My Focus: Ethical & Transparent AI/ML

My work encompasses advanced machine learning and Responsible AI. My technical expertise spans Machine learning modelling, Natural Language Processing, application of LLMs, and operationalization of Responsible AI principles. My approach involves embedding Fairness, Accountability, Transparency, and Reliability directly into the ML lifecycle, ensuring that systems are not only high-performing but also safe, compliant, and aligned with organizational and regulatory expectations.

I am driven by the challenge of turning ethical AI principles into practical workflows that scale. Whether I am building cloud-ready architectures, leading NLP initiatives with LLMs, designing model monitoring frameworks, or partnering with governance and risk teams, I aim to ensure every solution is transparent, explainable, and built for impact.

Key Skills & Tools

Data Science & AI: Machine Learning, Statistical Modelling, Anomaly Detection, A/B Testing, Natural Language Processing, Transformers, Python, SQL, Spark, MLOps
Responsible AI: Fairlearn, SHAP, LIME
Platforms: AzureML, MLflow, Databricks,
Data Viz: Power BI, Matplotlib for explainability dashboards
Project Management: Git, Version Controlling, Azure DevOps, Strategic Planning, Process & Resource Management, Agile/Scrum framework

My Structured Approach to Responsible AI Implementation

1. Robust De-Identification & Privacy Safeguards

Systematic removal, masking, or transformation of personally identifiable information (PII) throughout the data lifecycle. This ensures models are trained and evaluated on privacy-preserving datasets, maintaining regulatory compliance and protecting user confidentiality without compromising model performance.

2. Proactive Bias Audit

Rigorous pre-training assessment of data sources and features. Focus on identifying and mitigating systemic bias early, long before the model sees production data.

3. XAI Implementation

Utilizing SHAP and LIME for local and global model interpretability. This ensures full transparency and supports the "right to explanation" for all stakeholders.

4. Continuous Fairness Testing

Establishing automated pipelines to monitor for disparate impact and equity metrics across sensitive groups post-deployment, allowing for timely remediation.

✍️ Read My Latest Thoughts

Find my articles, technical guides, and commentary on Data Science, AI Ethics and Governance.

Visit My Blog

🔬 Featured Projects

Explainable AI for Credit Risk Decisions

HEALTHCARE | DE-IDENTIFICATION

The Challenge: A healthcare organization needed to safely leverage case notes for analytics and model development, but the raw text contained extensive PHI/PII, including names, dates, locations, and medical identifiers. Manual redaction was impossible and not scalable due to the volume of the reports, creating compliance risks under HIPAA and PHIPA, and blocking downstream AI/ML initiatives.

The Solution: I built an automated NLP-driven de-identification pipeline combining regex patterning, Named Entity Recognition (NER) using transformer models, and context-aware scoring to accurately detect and mask sensitive entities. The pipeline included a two-stage hybrid approach: deterministic rules for structured PHI and ML models for contextual PHI. To ensure enterprise readiness, I integrated confidence-threshold tuning, audit logs, and customizable redaction policies for different data-sharing use cases.

The Impact: Delivered 97%+ entity-level recall across 12 PHI categories while reducing false positives. Enabled safe, compliant data access and cut de-identification time from weeks to minutes, unlocking faster model development cycles and secure cross-team data sharing.

Explainable AI for Credit Risk Decisions

FINTECH | XAI

The Challenge: The existing credit decisioning workflow relied on a legacy scorecard with limited predictive power and no explainability framework, creating compliance friction and inconsistent lending decisions. As the product scaled, rising default rates signaled the need for a more accurate and transparent model to support digital loan applications.

The Solution: I managed the end-to-end development of a modern credit scoring model, replacing the legacy system with an interpretable architecture using Generalized Additive Models (GAMs). To strengthen transparency, I integrated SHAP-based explanations directly into the decision pipeline, enabling compliance-ready adverse action reasons and clearer customer communication. The full workflow included feature engineering, bias and drift monitoring, and governance controls to meet regulatory standards.

The Impact: The new credit scoring system reduced non-performing loans by 8.5% in the first year, improved decision consistency, and lowered manual review workloads. It also introduced a clear, defensible reasoning framework that strengthened trust across risk, compliance, and product teams.