ds

Extract and categorize learnings from completed experiments into docs/ds/learnings/ for future retrieval

Profile a dataset for structure, quality, distributions, and anomalies, then output an EDA report

Design an ML experiment with hypothesis, split strategy, leakage check, and evaluation plan

Frame a data science problem and plan the approach, surfacing relevant past learnings

Clean, validate, and transform raw data using automated preprocessing pipelines

Peer review an ML experiment for methodology, leakage, reproducibility, and statistical validity

Assess deployment readiness of a trained model and generate model card and deployment documentation

Run data quality validation using formal expectation suites, dbt tests, or data contracts

Profile datasets: missing rates, distributions, outliers, type issues. Use after loading data to characterize it before modeling.

Generate candidate features, check for leakage, and produce a feature registry. Use when building or evaluating feature sets.

Assess raw data quality and design preprocessing pipelines. Use when /ds:preprocess needs to determine what cleaning, validation, and transformation steps to apply.

Translate business questions into DS problems with target variables, metrics, and constraints. Use when starting a project or when the objective needs sharpening.

Define hypothesis, variables, split strategy, baselines, and comparison protocol. Use before running an experiment to lock down methodology.

Compute metrics, slice by subgroups, check calibration, and flag fairness gaps. Use after training to decide ship/iterate/abandon.

Evaluate whether an ML model is ready for production deployment by checking infrastructure, monitoring, rollback, and operational requirements. Use before shipping a model to production.

Extract reusable insights from experiment results and write them as searchable learning documents. Use at project end to capture what worked, failed, and surprised.

Audit ML experiments for reproducibility by checking seeds, versions, data hashes, and environment capture. Use when reviewing experiments before shipping or compounding.

Aeon API patterns for time series machine learning -- classification, regression, clustering, anomaly detection, segmentation, and similarity search. Use when /ds:experiment needs time-series-specific ML algorithms (ROCKET, InceptionTime, DTW classifiers), or /ds:eda needs temporal feature extraction (Catch22, ROCKET features) or change point detection. For classical statistical forecasting (ARIMA/SARIMAX) use statsmodels; for tabular ML pipelines use scikit-learn; for visualization use matplotlib.

Pre-model data preparation pipelines for cleaning, validation, transformation, and ETL orchestration. Use when raw data needs deduplication, schema validation, format conversion, or quality assurance before EDA or modeling.

Data quality validation with Great Expectations, dbt tests, and data contracts. Use when building formal validation rules, expectation suites, or data contracts for repeatable quality gates.

Systematic exploratory data analysis checklist covering structure, quality, distributions, relationships, and target analysis. Use when starting EDA on any dataset.

Standard format for logging ML experiments including hypothesis, config, results, and learnings. Use when running experiments to maintain a consistent record.

Detect file types and perform format-specific EDA across 200+ scientific data formats. Use when /ds:eda encounters non-tabular or unfamiliar data files, or when format-specific analysis guidance is needed.

Matplotlib API patterns for creating publication-quality visualizations. Use when /ds:eda needs distribution plots, correlation heatmaps, or relationship visualizations, or when /ds:experiment needs result plots (learning curves, confusion matrices, forecast visualizations). For standard ML diagnostic plots use scikit-learn display utilities; for statsmodels diagnostic plots use statsmodels built-in plotting; for quick statistical plots prefer seaborn.

Generate standardized model documentation following HuggingFace Model Card and NVIDIA Model Card++ formats. Use when preparing a model for deployment or handoff.

Pandas API patterns for DataFrame operations, data cleaning, aggregation, merging, and performance optimization. Use when generating pandas code for data loading, manipulation, or profiling in /ds:eda, /ds:preprocess, or /ds:experiment.

Polars expression API for high-performance DataFrame operations, lazy evaluation, joins, aggregations, and I/O. Use as a parallel alternative to pandas-pro when working with large datasets or generating Polars code for data loading, manipulation, or profiling in /ds:eda, /ds:preprocess, or /ds:experiment.

Verify that an ML experiment meets reproducibility requirements: random seeds, library versions, data hashes, environment capture. Use when reviewing experiments before shipping.

Scikit-learn API patterns for preprocessing, pipelines, model selection, and evaluation. Use when /ds:experiment needs to build sklearn pipelines, tune hyperparameters, or evaluate models.

Check Python environment for required DS/ML libraries and report versions or missing packages. Use when setting up a new project or debugging import errors.

SHAP API patterns for model interpretability -- explainer selection, feature attribution, and visualization. Use when /ds:experiment needs per-prediction explanations, global feature importance, or interaction analysis. For built-in tree importance and permutation importance use scikit-learn; for coefficient-based interpretation use statsmodels.

Select and implement appropriate train/validation/test split strategies based on data characteristics. Use when designing the evaluation framework for a model.

Guided statistical analysis with test selection, assumption checking, power analysis, and APA reporting. Use when /ds:experiment needs to design comparison protocols, validate assumptions, or report results.

Statsmodels API patterns for OLS, GLM, discrete choice, time series (ARIMA/SARIMAX), and diagnostics. Use when /ds:experiment needs statsmodels model fitting, diagnostics, or time-series forecasting, or /ds:eda needs VIF and stationarity checks. For guided test selection and APA reporting use statistical-analysis.

Detect target leakage in feature sets by checking temporal validity, feature-target correlation, and information flow. Use before training any model.

Hyperparameter tuning workflow reference -- strategy selection, Bayesian optimization with Optuna, search space design, and result analysis. Use when /ds:experiment needs to choose a tuning strategy, design search spaces, or analyze tuning runs.