Emily2040/data-science-agent-skills

ai-data-readiness-triage

Assess whether a dataset or data domain is ready for AI, analytics, or modeling before committing project budget. Use when triaging table schemas, missingness, label coverage, ownership, freshness, and PII risk to issue a go, fix-first, or stop decision.

Langue du texte source : anglais

baseline-ablation-lab

Design honest naive, rule-based, and linear baselines alongside feature and component ablations before accepting complex ML models. Use when challenging model complexity against latency, cost, seed stability, and minimum required lift.

Langue du texte source : anglais

causal-assumption-scout

Audit whether a data science claim requires causal identification rather than predictive correlation. Use when mapping treatments, outcomes, confounders, colliders, post-treatment controls, and quasi-experimental designs before claiming impact.

Langue du texte source : anglais

cost-aware-compute-planner

Estimate and gate GPU, cloud, and local compute budgets for data science experiments, training runs, and batch or online inference. Use when sizing 10% pilot stages, retry overhead, spot risk, and stop-loss thresholds before launching full-scale runs.

Langue du texte source : anglais

data-question-framing

Translate ambiguous business, product, policy, or research asks into measurable data science decision canvases. Use when locking down the decision owner, action threshold, analysis unit, target definition, time horizon, and guardrail metrics.

Langue du texte source : anglais

data-science-incident-rca

Run root-cause analysis for broken ML models, metric drops, pipeline failures, and corrupted dashboards. Use when correlating anomaly timestamps across upstream data, schema changes, code deploys, and serving layers to produce a countermeasure memo.

Langue du texte source : anglais

dataset-contract-forensics

Infer explicit data contracts, primary-key uniqueness rules, value ranges, and freshness SLAs from messy CSV, TSV, or warehouse extracts. Use when turning undocumented upstream tables into enforceable schema and validation tests.

Langue du texte source : anglais

domain-expert-collaboration-loop

Structure subject-matter expert (SME) annotation, rubric calibration, and edge-case adjudication for data science workflows. Use when measuring inter-rater agreement (Cohen's kappa), resolving label ambiguity, and encoding domain constraints.

Langue du texte source : anglais

drift-monitor-designer

Design production feature, prediction, and label drift monitors with calibrated alert thresholds and named on-call owners. Use when comparing baseline vs current distributions for mean shift, variance explosion, missingness spikes, and delayed ground truth.

Langue du texte source : anglais

evaluation-reality-check

Audit offline model evaluations for temporal split fidelity, subgroup slice disparity, calibration error, and metric-to-business alignment. Use when stress-testing benchmark numbers before approving deployment or executive reporting.

Langue du texte source : anglais

experiment-provenance-ledger

Capture cryptographic dataset hashes, git commit state, environment manifests, random seeds, and decision history for ML experiments. Use when building an auditable provenance ledger so model results can be reproduced or traced back to exact inputs.

Langue du texte source : anglais

feature-store-readiness

Evaluate whether an ML workflow needs a feature store and audit online-versus-offline feature parity. Use when checking point-in-time join correctness, serving latency SLAs, backfill cost, and entity-level value skew across training and serving.

Langue du texte source : anglais

leakage-adversary

Hunt for target, temporal, entity-group, preprocessing, and join leakage across train, validation, and test splits. Use when probing datasets and pipelines for split key overlap, post-outcome features, or suspiciously high validation scores.

Langue du texte source : anglais

messy-data-cleanroom

Design reversible, auditable data-cleaning pipelines that quarantine corrupt records without mutating raw source extracts. Use when standardizing schemas, deduplicating entities, and logging row-level exclusion impact.

Langue du texte source : anglais

metric-design-review-board

Stress-test primary, guardrail, and diagnostic metrics for denominator shifts, Goodhart gaming vectors, and decision alignment. Use when defining experiment or product KPIs so optimization cannot silently degrade user experience or unit economics.

Langue du texte source : anglais

missingness-mechanism-investigator

Diagnose whether missing data is MCAR, MAR, MNAR, or structurally absent before applying imputation or row deletion. Use when profiling column missingness rates, cross-tabulating missingness against outcomes, and choosing safe handling rules.

Langue du texte source : anglais

model-risk-compliance-memo

Produce governance-ready model risk assessments covering blast radius, affected populations, fairness slices, fallback policies, and rollback owners. Use when preparing high-impact ML or AI systems for compliance, legal, or release gate review.

Langue du texte source : anglais

notebook-repro-packager

Audit and package Jupyter notebooks for deterministic top-to-bottom execution across machines. Use when detecting out-of-order cell execution, hardcoded local paths, unseeded randomness, inline shell installs, or leaked secrets.

Langue du texte source : anglais

notebook-to-pipeline-surgeon

Refactor exploratory Jupyter notebooks into modular, testable Python pipeline stages with explicit input/output contracts. Use when extracting data loading, feature engineering, training, and evaluation out of stateful notebook cells.

Langue du texte source : anglais

oss-ds-supply-chain-auditor

Audit data science dependency manifests, pre-trained model weights, datasets, and licenses for security and reproducibility risks. Use when scanning requirements.txt, pyproject.toml, or environment files for unpinned packages and direct URL installs.

Langue du texte source : anglais

privacy-preserving-ds

Assess PII exposure, re-identification risk, and quasi-identifier linkage in analytics and ML datasets. Use when designing redaction, aggregation, k-anonymity, differential privacy, or clean-room controls before sharing or modeling sensitive data.

Langue du texte source : anglais

production-ml-handoff

Package prototype ML models into production engineering handoff bundles with input/output schemas, latency budgets, fallback baselines, and rollback triggers. Use when transitioning a validated model from data science exploration to ML engineering ownership.

Langue du texte source : anglais

semantic-lineage-mapper

Reconstruct end-to-end table, file, column, and metric lineage across SQL queries, Python scripts, notebooks, and BI configs. Use when tracing upstream sources and downstream blast radius before refactoring pipelines or deprecating fields.

Langue du texte source : anglais

small-data-power-planner

Calculate minimum detectable effect (MDE), sample-size requirements, and events-per-variable ratios for small or rare-event datasets. Use when preventing underpowered experiments and overfitted models on scarce data.

Langue du texte source : anglais

stakeholder-insight-storyteller

Translate technical data science findings into decision-grade executive memos with explicit asks, uncertainty intervals, and visible caveats. Use when converting notebooks, ablation tables, or experiment readouts into stakeholder recommendations.

Langue du texte source : anglais