Curated Projects

This page highlights a selective portfolio of projects that have been reviewed for provenance, methodology, reproducibility, data boundaries, and claim quality. Older notebooks and course exercises are not promoted simply because they exist.

Each featured project links to the maintained reconstruction in the DataForSolution curated portfolio, where the source, tests, audit notes, dependencies, and limitations are documented.

Healthcare & Medical Imaging

PETQuant Reliability

PET quantitative-ML reconstruction focused on small-sample limits, data-quality contracts, and defensible validation claims.

  • Python
  • PET safety
  • Validation

Chest CT Classification

Medical-imaging transfer-learning reconstruction emphasizing overfitting analysis, split integrity, and deterministic evaluation.

  • TensorFlow
  • ResNet50
  • Medical imaging

Pima Diabetes ML Evaluation

Leakage-safe SVM/MLP retrospective with clinically relevant missing-measurement handling and sensitivity/specificity reporting.

  • scikit-learn
  • Clinical metrics
  • Leakage control

Responsible & Explainable AI

SHAP + LIME Explainability

Class-aligned explanation workflow with corrected label ordering, scaled modeling, and defensible attribution aggregation.

  • SHAP
  • LIME
  • Explainability

Fairness Evaluation

Explicit group-fairness metrics and threshold analysis with fixed reference labels and documented AIF360 provenance.

  • AIF360
  • Group fairness
  • Thresholds

Adversarial Robustness

Bounded PGD and defense-evaluation reconstruction with preprocessing, threat-model, and denominator corrections.

  • PyTorch
  • PGD
  • Threat models

Applied Machine Learning & NLP

CIFAR-10 Generated-Image Analysis

Classifier-response diagnostics for generated images with a clear separation between classifier confidence and perceptual realism.

  • NumPy
  • Entropy
  • Generative AI

Data Engineering

Why only these projects?

The broader archive contains tutorials, course templates, duplicate notebooks, experiments with untraceable data, and projects whose original methodology does not support a strong public claim. Those materials are retained only when useful as historical evidence; they are not presented as equivalent to the maintained portfolio above.