claude-flow

{ "name": "MLE-STAR Machine Learning Engineering", "description": "Automated ML pipeline development using Search and Targeted Refinement methodology", "version": "1.0", "variables": { "dataset": "${DATASET_PATH}", "target": "${TARGET_COLUMN}", "output_dir": "./mle-star-output", "search_iterations": 3, "refinement_iterations": 5, "ensemble_size": 5 }, "agents": [ { "id": "search-coordinator", "type": "coordinator", "name": "Web Search Coordinator", "capabilities": ["web-search", "knowledge-synthesis"] }, { "id": "ml-researcher", "type": "researcher", "name": "ML Model Researcher", "capabilities": ["model-search", "paper-analysis", "kaggle-solutions"] }, { "id": "data-analyst", "type": "analyst", "name": "Data Analysis Expert", "capabilities": ["eda", "feature-analysis", "data-quality"] }, { "id": "ml-engineer-1", "type": "implementer", "name": "ML Engineer Alpha", "capabilities": ["model-implementation", "feature-engineering", "optimization"] }, { "id": "ml-engineer-2", "type": "implementer", "name": "ML Engineer Beta", "capabilities": ["model-implementation", "ensemble-methods", "hyperparameter-tuning"] }, { "id": "ablation-analyst", "type": "analyst", "name": "Ablation Study Analyst", "capabilities": ["performance-analysis", "component-testing", "impact-measurement"] }, { "id": "ensemble-architect", "type": "coordinator", "name": "Ensemble Strategy Architect", "capabilities": ["ensemble-design", "model-combination", "performance-optimization"] }, { "id": "validator", "type": "tester", "name": "Robustness Validator", "capabilities": ["data-leakage-detection", "validation", "debugging"] } ], "tasks": [ { "id": "web-search", "name": "Web Search for ML Solutions", "type": "research", "description": "Search for state-of-the-art models and successful approaches", "assignTo": ["search-coordinator", "ml-researcher"], "parallel": true, "claudePrompt": "Search for effective ML models and approaches for ${target} prediction using ${dataset}. Focus on: 1) State-of-the-art models, 2) Kaggle competition solutions, 3) Research papers with implementations, 4) Pre-trained models. Provide a comprehensive list of approaches with their key characteristics.", "input": { "dataset": "${dataset}", "target": "${target}", "iterations": "${search_iterations}" }, "output": { "models_found": "array", "approaches": "array", "initial_pipeline": "object" } }, { "id": "data-analysis", "name": "Comprehensive Data Analysis", "type": "analysis", "description": "Analyze dataset characteristics and requirements", "assignTo": "data-analyst", "depends": [], "claudePrompt": "Perform comprehensive EDA on ${dataset} for predicting ${target}. Analyze: 1) Data types and distributions, 2) Missing values and outliers, 3) Feature correlations, 4) Target variable characteristics, 5) Potential feature engineering opportunities. Provide actionable insights for ML pipeline design.", "input": { "dataset": "${dataset}", "target": "${target}" }, "output": { "data_insights": "object", "feature_recommendations": "array", "preprocessing_requirements": "array" } }, { "id": "initial-pipeline", "name": "Build Initial ML Pipeline", "type": "implementation", "description": "Create baseline pipeline using search results", "assignTo": "ml-engineer-1", "depends": ["web-search", "data-analysis"], "claudePrompt": "Build an initial ML pipeline incorporating the best approaches from web search. Use insights from data analysis to guide preprocessing and feature engineering. Create a modular pipeline with clear components: 1) Data preprocessing, 2) Feature engineering, 3) Model architecture, 4) Training strategy, 5) Evaluation metrics.", "input": { "search_results": "${web-search.output}", "data_insights": "${data-analysis.output}", "dataset": "${dataset}", "target": "${target}" }, "output": { "pipeline_code": "string", "baseline_score": "number", "components": "object" } }, { "id": "ablation-study", "name": "Component Ablation Analysis", "type": "analysis", "description": "Identify critical pipeline components through ablation", "assignTo": "ablation-analyst", "depends": ["initial-pipeline"], "iterative": true, "maxIterations": "${refinement_iterations}", "claudePrompt": "Conduct systematic ablation study on the ML pipeline. For each component (preprocessing, feature engineering, model, postprocessing): 1) Remove/disable the component, 2) Measure performance impact, 3) Identify the most critical component. Provide detailed analysis of which component has the highest impact on performance.", "input": { "pipeline": "${initial-pipeline.output.pipeline_code}", "baseline_score": "${initial-pipeline.output.baseline_score}" }, "output": { "critical_component": "string", "impact_scores": "object", "ablation_results": "array" } }, { "id": "targeted-refinement", "name": "Focused Component Optimization", "type": "optimization", "description": "Deep refinement of identified critical component", "assignTo": ["ml-engineer-1", "ml-engineer-2"], "depends": ["ablation-study"], "parallel": true, "iterative": true, "claudePrompt": "Focus on optimizing the ${ablation-study.output.critical_component} component. Generate 10+ variations: 1) Different implementation approaches, 2) Hyperparameter variations, 3) Alternative algorithms, 4) Novel techniques. Test each variation and track performance. Select the best performing variation.", "input": { "critical_component": "${ablation-study.output.critical_component}", "current_pipeline": "${initial-pipeline.output.pipeline_code}", "target_score": "${initial-pipeline.output.baseline_score}" }, "output": { "best_variation": "object", "improved_pipeline": "string", "new_score": "number", "all_variations": "array" } }, { "id": "ensemble-creation", "name": "Advanced Ensemble Strategy", "type": "ensemble", "description": "Create sophisticated ensemble from best models", "assignTo": "ensemble-architect", "depends": ["targeted-refinement"], "claudePrompt": "Design an advanced ensemble strategy using the top ${ensemble_size} model variations. Consider: 1) Model diversity and complementarity, 2) Weighted voting vs stacking vs blending, 3) Cross-validation for weight optimization, 4) Out-of-fold predictions. Implement the ensemble and optimize for maximum performance.", "input": { "candidate_models": "${targeted-refinement.output.all_variations}", "ensemble_size": "${ensemble_size}", "validation_data": "${dataset}" }, "output": { "ensemble_strategy": "object", "ensemble_code": "string", "ensemble_score": "number", "model_weights": "array" } }, { "id": "robustness-validation", "name": "Robustness and Quality Checks", "type": "validation", "description": "Comprehensive validation and debugging", "assignTo": "validator", "depends": ["ensemble-creation"], "parallel": true, "claudePrompt": "Perform comprehensive robustness validation: 1) Check for data leakage in all pipeline stages, 2) Verify train/test separation integrity, 3) Debug any execution errors, 4) Validate reproducibility, 5) Check resource usage and optimization opportunities. Fix any issues found and ensure production readiness.", "input": { "final_pipeline": "${ensemble-creation.output.ensemble_code}", "dataset": "${dataset}", "all_components": "${initial-pipeline.output.components}" }, "output": { "validation_report": "object", "issues_found": "array", "fixes_applied": "array", "production_ready": "boolean" } }, { "id": "deployment-package", "name": "Create Deployment Package", "type": "deployment", "description": "Package final solution for deployment", "assignTo": "ml-engineer-1", "depends": ["robustness-validation"], "claudePrompt": "Create a production-ready deployment package: 1) Clean and optimize code, 2) Create requirements.txt, 3) Add comprehensive documentation, 4) Include model serialization, 5) Create inference API, 6) Add monitoring hooks. Package everything in ${output_dir}.", "input": { "validated_pipeline": "${robustness-validation.output}", "output_dir": "${output_dir}" }, "output": { "deployment_path": "string", "api_endpoint": "string", "documentation": "string", "performance_metrics": "object" } } ], "hooks": { "pre-task": "npx claude-flow@alpha hooks pre-task --description \"${task.name}\" --auto-spawn-agents false", "post-task": "npx claude-flow@alpha hooks post-task --task-id \"${task.id}\" --analyze-performance true", "on-error": "npx claude-flow@alpha hooks notify --message \"Error in ${task.id}: ${error.message}\" --telemetry true" }, "settings": { "maxConcurrency": 4, "timeout": 3600000, "retryPolicy": "exponential", "failurePolicy": "continue", "memoryCoordination": true, "progressTracking": "detailed", "outputFormat": "stream-json" } }