mlops-stack

Anti-slop agentic engineering co-pilot. Teaches the Research-Plan-Implement (RPI) workflow, context management, quality gates, per-agent isolation, and anti-slop patterns for building software with AI coding agents. Produces agent-workflow.md or project configuration files. Part of the mlops-tabular skill family but independently invocable for any software project.

Full software engineering and ML-specific code review co-pilot. Reviews Python code for quality, security, testing, type safety, and ML-specific issues including data leakage, training-serving skew, feature engineering smells, and reproducibility. Produces structured review findings by severity. Part of the mlops-tabular skill family. Invoke via /mlops-tabular or directly for any Python/ML code review.

System design co-pilot covering both general distributed systems and ML-specific infrastructure. Guides users through API design, database design, scalability, reliability, ML serving patterns, feature stores, training pipelines, and ML platform architecture. Produces system_design.md. Part of the mlops-tabular skill family. Invoke via /mlops-tabular or directly for any system design problem.

Production-grade MLOps co-pilot for tabular data. Guides users end-to-end from business problem through system design, implementation, deployment, and monitoring. Adapts dynamically to the user's specific problem, dataset, constraints, and chosen orchestration framework. Use when asked to build an ML product on tabular data, productionize a model, set up MLOps infrastructure, or when users describe a business problem they want to solve with machine learning on structured data. Proactively invoke when: user describes a business problem solvable with tabular ML, mentions prediction/classification/regression on structured data, or asks about MLOps best practices for a specific project.

Deep-dive MLOps architecture design for tabular data. Walks through all 9 sub-phases of system design: full pipeline explanation (10 stages, 5 pipelines, maturity levels), data plan, feature plan, training plan, deployment plan, monitoring plan, versioning plan, ZenML stack selection, and architecture document production. Reads problem_statement.md, produces architecture.md. Part of the mlops-tabular skill family.

Deep-dive data foundation and feature engineering for tabular ML. Covers project setup, data loading with validation, EDA, and preprocessing (null handling, scaling with formulas, categorical encoding with target encoding smoothing, training-serving skew prevention with sklearn.Pipeline). Reads problem_statement.md and architecture.md. Part of the mlops-tabular skill family.

Deep-dive deployment, monitoring, and production hardening for tabular ML. Covers drift detection (data vs concept drift, KS/Chi-squared/PSI/Wasserstein with thresholds), deployment strategies (shadow/canary/blue-green/A-B), four-layer monitoring ladder, incident response, feedback loop dangers, production hardening, and shipping. Part of the mlops-tabular skill family.

Deep-dive problem framing for tabular ML. Guides users through the six-word ML suitability test, three legitimate paths (Build ML / Rules / Not Now), problem statement template, metric ladder, seven discovery questions, and six forcing questions. Produces problem_statement.md. Part of the mlops-tabular skill family. Invoke via /mlops-tabular or directly when you need focused problem framing.

Deep-dive model training and evaluation for tabular ML. Covers experiment tracking with four reproducibility elements, baseline models as mandatory, evaluation with slice-level analysis and confidence intervals via bootstrapping, and class imbalance handling with the four-factor decision framework. Part of the mlops-tabular skill family.