ai-research-pipeline

End-to-end applied AI/ML research pipeline. Takes a research question and dataset, runs literature review, multi-method classification/regression analysis with parallel model tracks, and produces a complete manuscript (Markdown + LaTeX/PDF). Use when user says "application pipeline", "applied analysis", "analyze my data and write paper", "end-to-end analysis", or wants to go from raw data to manuscript. Covers all data modalities: NLP text, structured/tabular, and image data.

Server-side extension that completes the full analysis pipeline for image classification after vera-ai-image-reviewing has run. Adds ResNet50, EfficientNet-B0, VGG16, DenseNet121 CNN classifiers with transfer learning (feature extraction and full fine-tuning), subgroup analysis by metadata or image properties, advanced models (ViT transformer with optional tabular fusion and hyperparameter search), ensemble methods, cross-method comparison with unified feature attribution (GradCAM, attention maps) on a 0-100 scale, and manuscript-ready methods.md and results.md. Applies output variation and code style diversity for natural, non-repetitive output. Open-source skill. Triggered after vera-ai-image-reviewing completes and its PART 0–2 artifacts are present (see ../../CROSS-SKILL-INTERFACE.md). If invoked directly without those artifacts, halts and prompts the user to run testing first or supply equivalent PART 0–2 code.

Runs data quality diagnostics and baseline image classification. Produces class distribution tables, sample image grids, image size and channel statistics, a baseline CNN or pre-trained feature extractor with Logistic Regression, weighted F1 and macro AUC (bootstrapped 95% CIs), confusion matrix, and ROC curves. Ends with a recommendation block listing additional models available in the analysis workflow. Outputs Python scripts with 2 publication-quality plots. Triggered when user has image data and says "image classification," "computer vision," "CNN," "object recognition," "medical imaging," "image detection," "visual recognition," "photo classification," "X-ray classification," or describes a task involving classifying images into categories. Does not handle free-text NLP or tabular-only data.

End-to-end AI/ML methodology research pipeline. From research direction to publication-ready manuscript with novel architectures, training strategies, or evaluation methods. Includes idea discovery, implementation (model code, ablation studies, benchmark experiments), external review through the configured reviewer bridge, and paper writing (LaTeX + PDF). Use when user says "methodology pipeline", "develop new method", "research pipeline", "full pipeline", "run everything", or wants the complete autonomous AI/ML methodology research workflow. Designed for overnight autonomous execution.

Server-side extension that completes the full analysis pipeline for NLP text classification after vera-ai-nlp-reviewing has run. Adds SVM, Random Forest, LightGBM classifiers with TF-IDF and optional extra features, subgroup analysis by metadata or text properties, deep learning models (GRU, TextCNN, ALBERT with optional tabular fusion and hyperparameter search), cross-method comparison with unified feature importance on a 0-100 scale, and manuscript-ready methods.md and results.md. Applies output variation and code style diversity for natural, non-repetitive output. Open-source skill. Triggered after vera-ai-nlp-reviewing completes and its PART 0–2 artifacts are present (see ../../CROSS-SKILL-INTERFACE.md). If invoked directly without those artifacts, halts and prompts the user to run testing first or supply equivalent PART 0–2 code.

Runs class balance diagnostics and baseline text classification for NLP tasks. Produces class distribution tables, text length statistics, vocabulary analysis, TF-IDF feature inspection, a baseline Logistic Regression classifier with weighted F1 and macro AUC (bootstrapped 95% CIs), confusion matrix, and ROC curves. Ends with a recommendation block listing additional models and analyses available in the analysis workflow. Outputs Python scripts with 2 publication-quality plots. Triggered when user has text data and says "text classification," "NLP," "sentiment analysis," "AI detection," "language model detection," "document classification," "spam detection," "text mining," "topic classification," or describes a task involving classifying text into categories. Does not handle structured/tabular-only data or image data.

Server-side extension that completes the full analysis pipeline for structured/tabular data after vera-ai-structured-reviewing has run. Adds SVM, Random Forest, XGBoost, LightGBM, CatBoost classifiers/regressors with missing value imputation, encoding, and scaling, subgroup analysis by metadata variables, deep learning models (MLP, TabNet with optional hyperparameter search), stacking ensemble with meta-learner, cross-method comparison with unified feature importance on a 0-100 scale, and manuscript-ready methods.md and results.md. Supports BOTH classification (F1/AUC) AND regression (RMSE/R2/MAE). Applies output variation and code style diversity for natural, non-repetitive output. Open-source skill. Triggered after vera-ai-structured-reviewing completes and its PART 0–2 artifacts are present (see ../../CROSS-SKILL-INTERFACE.md). If invoked directly without those artifacts, halts and prompts the user to run testing first or supply equivalent PART 0–2 code.

Runs data quality diagnostics and baseline classification/regression for structured (tabular) data. Produces missing value analysis, feature distributions, correlation matrix, class balance check, outlier detection, a baseline LightGBM classifier with weighted F1 and macro AUC (bootstrapped 95% CIs), feature importance, confusion matrix, and ROC curves. Ends with a recommendation block listing additional models available in the analysis workflow. Outputs Python scripts with 2 publication-quality plots. Triggered when user has tabular/structured data and says "tabular data," "structured data," "classification," "regression," "feature engineering," "predict from columns," "CSV classification," "spreadsheet," "predict outcome," or describes a task involving predicting from numeric/categorical columns. Does not handle free-text NLP or image data.