Data Scientist
realnghon/data-scientist
De dados bagunçados a relatórios com evidência em minutos
Ver no GitHubSkill para análise de dados estruturados que transforma CSVs/Excel/Parquet em relatórios com evidência estatística. Perfila dados, valida qualidade, seleciona método estatístico apropriado e gera gráficos com intervalos de confiança e limitações documentadas.
Instalação
npx convez-ai realnghon/data-scientist1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Análise de yield e variáveis de processo em manufatura
- Avaliação de testes A/B com verificação de múltiplas métricas
- Detecção de anomalias e decomposição de séries temporais
- Análise de causa-raiz com junção de múltiplas fontes
- Cálculo de índices de capacidade (Cp/Cpk) e controle estatístico
Veredito

Skill robusto e prático para quem precisa sair de análises manuais no Excel. Os 8 gates de qualidade e framework de evidência em 3 níveis evitam conclusões enganosas — essencial para manufatura e experimentação. Suporte multi-plataforma sólido (Claude Code, Cursor, Cline, etc.).
README
A tradução para o português deste README ainda está na fila.
What it does
Give it a CSV/Excel/Parquet file and a question. It profiles your data, checks if analysis is possible, picks the right statistical method, runs it, and produces a report with charts, confidence intervals, and limitations.
Built for:
- Manufacturing analytics (yield drivers, SPC, Cp/Cpk, defect rates)
- A/B test evaluation (multi-metric tradeoffs, assumption checks)
- Time series (seasonality, anomaly detection, change-points)
- Root cause analysis (multi-source joins, Simpson's paradox)
Key features:
- 8-dimension data quality check before analysis (blocks if data is insufficient)
- 3-tier evidence framework (separates reliable findings from directional signals)
- Manufacturing-grade methods (SPC, MSA, DOE, interaction effects)
- Validated on 3 comprehensive test cases with two-line scoring (process adherence + outcome quality)
Quick start
Install
# For Claude Code / Codex / Cursor
npx skills add realnghon/data-scientist
Supports Claude Code, Codex, Cursor, OpenCode, Cline, Windsurf, GitHub Copilot, Gemini CLI. See INSTALL.md for platform-specific setup.
Run
# Claude Code / Codex
/ds-analyze examples/manufacturing_yield/dataset.csv
# Or just ask
"Analyze examples/ab_test/dataset.csv — is treatment better than control?"
Output: Data profile → readiness check → method plan → execution → charts → report with limitations.
Use cases
Manufacturing:
- Yield driver ranking (which process variables affect output?)
- SPC control charts (is the line in control? Cpk calculation)
- Equipment anomaly detection (seasonal patterns, change-points)
Experiments:
- A/B test evaluation (multi-metric tradeoffs, SRM check, effect size + CI)
- DOE analysis (interaction effects, optimal parameter ranges)
Operations:
- Time series forecasting (decompose seasonality, detect anomalies)
- Root cause analysis (join multiple data sources, trace mechanisms)
What's included
- 1 skill — 42KB workflow with 8 reference documents and tested helpers
- 11 method families — regression, A/B tests, SPC, time series, correlation, survival analysis, etc.
- 8 data quality gates — sample size, missingness, grain, leakage, balance, measurement reliability
- 21 chart functions — ready-made plots for all analysis types (no hand-written matplotlib)
- 3 evaluation cases — manufacturing, business, time series (scored on process adherence + outcome quality)
How it works
intake → readiness check → data shaping → method selection → execution → critic → report
Gates that stop bad analysis:
- If data is too small / too many missing / wrong grain → emits data request, stops analysis
- If p ≥ 0.05 → downgrade to "directional signal", never claim reliable
- If spec limits look wrong → flag measurement issue before proceeding
3-tier evidence:
- Tier-1 (reliable): p < 0.05, cross-checked, effect size + CI, physically plausible
- Tier-2 (directional): p ≥ 0.05 or failed assumptions, signal present but uncertain
- Tier-3 (no evidence): tested but rejected, reported to show what was checked
Development
# Run tests
npm test
# Profile a dataset (no AI in the loop)
python plugins/data-scientist/skills/analysis-workflow/scripts/profile_dataset.py data.csv
# Install with Excel/Parquet support
pip install -e ".[io]"
Evaluation: 3 comprehensive test cases covering manufacturing (multi-table root cause), business analytics (A/B test with Simpson's paradox), and time series (anomaly + seasonality). Scored on two independent lines — process_score (deterministic workflow-adherence) and outcome_score (agent-judge conclusion quality) — compared as before/after distributions to drive skill iteration. See evals/ for methodology.
License
MIT — see LICENSE.
Contributing
Issues and PRs welcome. High-leverage contributions:
- New method entries (logistic regression, GLM, Bayesian methods)
- Golden templates for new domains (logistics, finance, web analytics)
- Platform integrations (JetBrains, VS Code extensions)