This repository is a curated portfolio of selected data science, machine learning, healthcare AI, responsible-AI, and data-engineering work.
It is not a raw archive of every notebook I have created. Older coursework and experiments are reviewed individually. A project is promoted into portfolio/ only when its provenance, methodology, reproducibility, and claims are strong enough to present publicly.
Each reconstructed project aims to:
- preserve the historical learning context without overstating authorship;
- distinguish original work from course, tutorial, framework, or starter code;
- identify and document methodological defects found during audit;
- use reproducible environments and testable source modules instead of notebook state alone;
- keep datasets and external artifacts traceable to documented sources/licenses;
- avoid publishing credentials, private data, PHI, or unsupported claims;
- report limitations as clearly as results.
- PETQuant Reliability — reconstruction of a PET quantitative-ML workflow with explicit small-sample, data-quality, and validation boundaries.
- Chest CT Classification — transfer-learning reconstruction focused on evaluation discipline, overfitting, split integrity, and deterministic inference.
- Pima Diabetes ML Evaluation — leakage-safe SVM/MLP retrospective with clinically relevant missing-measurement handling and sensitivity/specificity reporting.
- SHAP + LIME Explainability — class-aligned explainability workflow with corrected label ordering and attribution aggregation.
- Fairness Evaluation — explicit group-fairness metrics and threshold analysis with fixed reference labels and documented AIF360 provenance.
- Adversarial Robustness — bounded PGD and defense-evaluation reconstruction with preprocessing, threat-model, and denominator corrections.
- Automobile Insurance Fraud Classification — temporal, imbalance-aware fraud evaluation showing why high accuracy can coexist with zero minority-class recall.
- Restaurant Review Sentiment Evaluation — leakage-safe NLP model comparison using fold-local TF-IDF pipelines and a licensed UCI Yelp data source.
- CIFAR-10 Generated-Image Analysis — classifier-response diagnostics for generated images with a clear distinction between model confidence and perceptual realism.
- AWS Glue + EMR PySpark Ingestion Validation — catalog-resolution and ingestion-quality reconstruction based on a historical embedded-header defect discovered in an EMR/Glue run.
View the full portfolio index →
Historical notebooks, tutorials, duplicate copies, projects with untraceable datasets, very small experiments, and analyses whose methodology cannot support their original claims are not shown as featured projects. They may remain in archival storage or Git history when useful as learning/provenance evidence, but they are intentionally absent from the current portfolio surface.
The current branch is intentionally minimal: README.md plus the canonical portfolio/ directory. Legacy root notebooks and old website files have been removed from the current tree so the repository listing itself reflects the curated standard.
Each featured project maintains its own README, audit notes, source modules, tests, environment definition, and interpretation boundaries as appropriate.
Portfolio CI uses Python 3.11 to parse all Python source and notebook code cells, check local Markdown links and sensitive/local-path patterns, detect whitespace errors, and run the 76 network-free unit tests across all ten projects. The shared test environment keeps scikit-learn below 1.10, which also preserves Pima's required <1.11 bound until its probability-estimation methodology is deliberately migrated and revalidated.
CI deliberately does not execute notebooks, download datasets or models, install optional research frameworks, use GPUs, retrain models, or regenerate project results. Those operations are excluded because the repository's baseline checks protect code and presentation integrity without implying that historical experiments have been reproduced.
Several projects use healthcare or other high-stakes datasets. They are educational/research engineering projects and are not clinical, diagnostic, financial, or security decision systems unless a project explicitly states otherwise.