Interactive Data Science Lab

Interactive technical notes

Data science ideas you can manipulate.

Small interactive demonstrations that expose the mechanics behind statistical learning: thresholds, trade-offs, projections, uncertainty, and model behavior. They are compact technical notes for people who want to see how the concepts work.

Interactive Lab

Explore how ROC and Precision-Recall curves respond when prevalence, class separation, and decision thresholds change. The demo makes visible a common production ML risk: a model can look acceptable under one metric while failing the operating point that matters for the positive class.

ROC Precision-Recall Class imbalance Thresholds Model evaluation

Technical challenge: Choosing metrics and thresholds when false positives and false negatives have different operational costs.
My contribution: Built an interactive simulator that links score distributions, confusion-matrix behavior, and evaluation curves.
Industry transfer: Fraud detection, medical triage, rare-event monitoring, ranking, recommender evaluation.
Connected work: TOROS rare-event detection and production model evaluation.

Manipulate a two-dimensional data cloud and observe how principal components rotate, project, and reconstruct the data. The applet gives a concrete view of what dimensionality reduction preserves, what it discards, and why that matters before feeding compressed representations into downstream models.

PCA Linear algebra Dimensionality reduction Feature compression Reconstruction

Technical challenge: Reducing dimensionality without losing the structure needed for the task.
My contribution: Designed a visual explanation that connects covariance geometry to model-ready representations.
Industry transfer: Embeddings, denoising, exploratory analysis, feature engineering, scientific data reduction.
Connected work: VVV/VVVX feature extraction and DEEPz high-dimensional regression.

Adjust optimization settings and watch how the path over a loss surface changes. The demo shows why learning rate, curvature, and initialization affect convergence, instability, and training time in machine learning systems.

Optimization Gradient descent Numerical methods Loss functions Model training

Technical challenge: Making iterative optimization behavior visible enough to diagnose convergence failures.
My contribution: Built an interactive loss-surface view that connects parameter updates to model training dynamics.
Industry transfer: Deep learning, forecasting models, calibration, optimization-heavy analytics workflows.
Connected work: Production ML training workflows and statistical modeling.

Change model complexity and data conditions to see how training fit can diverge from generalization. The applet turns overfitting into an observable behavior rather than an abstract warning, making it useful for model selection and risk-aware validation.

Generalization Bias-variance Validation Model selection Regularization

Technical challenge: Distinguishing useful flexibility from unstable fitting under noisy data.
My contribution: Created an interactive simulator for comparing training behavior with out-of-sample performance.
Industry transfer: Experiment design, forecasting, predictive maintenance, risk models, applied ML validation.
Connected work: Mercado Libre experimentation and robust scientific inference.

Compare how classifiers partition the same feature space under different assumptions. The visualization helps explain why model choice is not only about accuracy: geometry, noise, class overlap, and inductive bias shape the decisions a system will make.

Classification Decision boundaries Supervised learning Model comparison Statistical learning

Technical challenge: Understanding how classifier assumptions translate into operational decisions.
My contribution: Built a visual comparison tool for linking model families to boundary geometry.
Industry transfer: Segmentation, risk scoring, customer modeling, medical classification, anomaly triage.
Connected work: Variable-star classification and production ML decision systems.

Visualize how regularization changes the feasible solutions and the models that emerge from data. The demo highlights the engineering role of regularization: not just improving metrics, but making models more stable, interpretable, and reliable under limited or noisy evidence.

Regularization L1/L2 penalties Model stability Feature selection Regression

Technical challenge: Preventing models from becoming too sensitive to noise or weak signals.
My contribution: Created an interactive view of how penalty geometry shapes learned parameters.
Industry transfer: Risk estimation, price prediction, forecasting, sparse modeling, production regression.
Connected work: DEEPz uncertainty-aware regression and ML model calibration.

Interactively fit a mixture model and observe how expectation-maximization alternates between assigning probability mass and updating component parameters. The applet makes soft clustering and uncertainty visible, which is central to segmentation problems where group membership is not obvious.

EM algorithm Mixture models Clustering Probabilistic modeling Uncertainty

Technical challenge: Recovering latent groups when the data are noisy, overlapping, or partially ambiguous.
My contribution: Built a two-dimensional simulator that exposes the iterative mechanics behind soft clustering.
Industry transfer: Customer segmentation, anomaly grouping, market discovery, spatial analytics.
Connected work: Cosmic structure analysis and unsupervised pattern discovery.

Explore how hierarchical clustering builds structure from pairwise distances and linkage rules. The applet is useful for understanding why clustering is not one result but a modeling decision that depends on scale, distance, noise, and the structure one wants to preserve.

Clustering Linkage methods Dendrograms Distance metrics Pattern discovery

Technical challenge: Identifying structure across scales without imposing a fixed number of groups too early.
My contribution: Built an interactive clustering view that connects local distances to global grouping decisions.
Industry transfer: Customer segmentation, document clustering, taxonomy design, operations analytics.
Connected work: Cosmic void detection and large-scale spatial structure analysis.

Adjust the kernel behavior and see how nonlinear decision regions form. The demo gives a practical view of the trade-off between flexibility and stability: highly adaptive boundaries can capture complex patterns, but can also become sensitive to noise.

SVM RBF kernel Nonlinear models Margins Classification

Technical challenge: Capturing nonlinear structure while preserving a stable decision rule.
My contribution: Built a visual tool for understanding how kernel parameters reshape the classifier boundary.
Industry transfer: Image classification, anomaly detection, quality control, risk models.
Connected work: Computer vision pipelines and automated object detection.

Compare individual noisy learners with their aggregate behavior. The applet shows the central intuition behind ensemble models: independent errors can partially cancel out, reducing variance and improving reliability without requiring every base model to be perfect.

Ensembles Random Forests Variance reduction Model stability Bootstrap intuition

Technical challenge: Improving prediction stability when individual models are sensitive to the data sample.
My contribution: Created an interactive demonstration of variance reduction through aggregation.
Industry transfer: Random Forest modeling, forecasting, ranking, noisy industrial prediction systems.
Connected work: Random Forest classification in TOROS candidate ranking.

Observe how small perturbations in the training data can change a decision tree. The applet captures a practical issue in interpretable modeling: simple rules can be easy to explain, but their apparent clarity may hide sensitivity to sampling noise.

Decision trees Instability Validation Robustness Interpretability

Technical challenge: Separating interpretable structure from unstable artifacts of a specific sample.
My contribution: Built a focused visualization of tree boundary sensitivity under small data changes.
Industry transfer: Credit policies, churn models, operational rules, interpretable ML governance.
Connected work: Model validation and production experimentation.

Explore how inter-rater agreement changes under different labeling patterns. The applet highlights a frequent ML bottleneck: model quality is bounded by the reliability of the data labels, and agreement metrics can reveal problems that raw counts hide.

Label quality Cohen's kappa Evaluation Annotation workflows Data quality

Technical challenge: Quantifying whether labels are consistent enough to support reliable model training and evaluation.
My contribution: Built a simulator for interpreting agreement metrics under changing base rates and disagreement patterns.
Industry transfer: Human-in-the-loop ML, annotation QA, medical review, content moderation, audit workflows.
Connected work: Human validation in TOROS and applied ML data quality work.

Why This Lab Exists

Explainability

Interactive models make assumptions, thresholds, and failure modes visible instead of hiding them behind summary metrics.

Technical judgment

The applets focus on decisions practitioners actually make: which metric to trust, what to compress, and what uncertainty remains.

Technical communication

They demonstrate the ability to mentor, document, and communicate complex ideas to technical teams and stakeholders.

Interactive Data Science Lab

Data science ideas you can manipulate.

Interactive Lab

ROC and Precision-Recall Explorer

PCA Projection Explorer

Gradient Descent Playground

Bias-Variance and Generalization

Decision Boundary Explorer

Regularization Explorer

Gaussian Mixture and EM Explorer

Hierarchical Clustering Lab

Kernel SVM Explorer

Ensemble Variance Explorer

Decision Tree Instability

Agreement and Label Quality Explorer

Why This Lab Exists

Explainability

Technical judgment

Technical communication