convez.ai
SkillDados & AnalyticsIntermediário
realnghon

Data Scientist

realnghon/data-scientist

De dados bagunçados a relatórios com evidência em minutos

Ver no GitHub
InicialPluginScore · de 0 a 100

Skill 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-scientist

1 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

Estrelas
2
Forks
0
Acompanhando
0
Saúde
57%
Último commit
há 4 meses
Versão
—
Issues abertas
0
Contribuidores
3

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:

  1. If data is too small / too many missing / wrong grain → emits data request, stops analysis
  2. If p ≥ 0.05 → downgrade to "directional signal", never claim reliable
  3. 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)
Ver o original no GitHub

Bom para

Relacionados