portfolio

Curated Portfolio

These projects were selected from a larger archive of notebooks, coursework, experiments, and research work. Each featured project has been reviewed for provenance, methodology, reproducibility, data boundaries, and claim quality before being promoted here.

Project index

Project	Domain	Core technologies	What the reconstruction demonstrates
PETQuant Reliability	Healthcare AI / PET	Python, NumPy, pytest	Small-data validation, data-quality contracts, conservative reliability gates
Chest CT Classification	Medical imaging	TensorFlow, scikit-learn, pytest	Transfer-learning discipline, overfitting analysis, deterministic validation/test handling
Pima Diabetes ML Evaluation	Healthcare ML	scikit-learn, pandas, OpenML	Leakage-safe preprocessing, missing-measurement handling, sensitivity/specificity
SHAP + LIME Explainability	Explainable AI	SHAP, LIME, scikit-learn	Stable class alignment, attribution aggregation, interpretation limits
Fairness Evaluation	Responsible AI	Python, AIF360, pytest	Group-fairness metrics, fixed-label threshold evaluation, utility tradeoffs
Adversarial Robustness	ML security	PyTorch, ART, pytest	Bounded PGD, threat-model discipline, clean-vs-robust utility
Automobile Insurance Fraud Classification	Imbalanced ML	scikit-learn, pandas, PR-AUC	Temporal holdout, threshold selection, minority-class tradeoffs
Restaurant Review Sentiment Evaluation	NLP	TF-IDF, scikit-learn, VADER	Fold-local text modeling and stratified out-of-fold comparison
CIFAR-10 Generated-Image Analysis	Generative AI evaluation	NumPy, classifier diagnostics, pytest	Probability validation, entropy/margin analysis, realism boundary
AWS Glue + EMR PySpark Ingestion Validation	Data engineering	PySpark, AWS Glue, EMR	Catalog resolution, embedded-header detection, fail-closed ingestion checks

Selection standard

A historical project is not promoted merely because it runs or once received a good grade. The reconstruction process asks:

1. Provenance — What was original work, what came from a tutorial/course/example, and can the distinction be documented?
2. Data legitimacy — Is the dataset traceable, appropriately licensed, de-identified where necessary, and suitable for the claim being made?
3. Methodological validity — Are train/validation/test boundaries, preprocessing, metrics, denominators, and evaluation procedures defensible?
4. Reproducibility — Can the important logic be expressed as source code with explicit dependencies and tests rather than relying on notebook state?
5. Portfolio value — Does the project demonstrate a distinct engineering or analytical skill that is not already represented better elsewhere?

Projects that fail this standard are archived, deferred, or used only as methodology-audit evidence.

Reconstructed project pattern

Most projects follow a common structure:

project/
  README.md
  pyproject.toml
  src/
  tests/
  docs/AUDIT.md
  notebooks/
  data/README.md   # when external data provenance is relevant

The AUDIT.md files are intentional. They document historical bugs, provenance boundaries, and interpretation limitations rather than rewriting history to make an older project look cleaner than it actually was.

Current portfolio coverage

The selected projects collectively demonstrate:

• healthcare and medical-imaging technology,
• classical and neural-network model evaluation,
• explainable and responsible AI,
• adversarial robustness,
• generative-model diagnostics,
• imbalanced classification,
• natural-language processing,
• AWS/PySpark data engineering,
• reproducible Python packaging and regression testing.

The portfolio will remain selective. Additional historical work will be added only when it contributes a distinct, defensible capability.