ML Pipeline Orchestrator Agent

• 

  Pipeline successfully executed page

• 

  Final confirmation of Job page (Successfully completed)

• 

  Pipeline running page

• 

  Creating a new issue page

• 

  Final report page - 1

• 

  Initial landing output page (Process has been started)

• 

  Issues history page

• 

  Agents running page (one after another)

• 

  Agents running page ( one after another)

• 

  Final report page - 2

• 

  Agents running page (one after another)

• 

  Official GitLab repo code base

• 

  VS code , code deployment/development page

• 

  Job description page on executing pipeline

Inspiration

Building ML pipelines manually takes 2-3 days. We wanted to automate the entire process with autonomous agents.

What it does

Transforms a GitLab issue into a production-ready ML pipeline in 2 minutes. File an issue with your dataset, and 5 autonomous agents orchestrate the entire pipeline creation — from data analysis to model training to results reporting.

How we built it

• Trigger Agent: Reads GitLab issues and extracts task specifications
• Dataset Analyst: Profiles CSV datasets and identifies key characteristics
  
• Strategist Agent: Claude Sonnet 4.6 reasons about optimal ML strategy
• Code Generator: Generates 8-file production ML pipeline (preprocess, train, evaluate, etc.)
• Reporter Agent: Trains model, evaluates performance, posts results to GitLab

Challenges we ran into

• Multi-agent orchestration and state management
• Ensuring Claude generates syntactically correct production code
• Integrating with GitLab CI/CD pipelines automatically
• Handling edge cases in dataset profiling

Accomplishments that we're proud of

✅ 5 autonomous agents working seamlessly together ✅ 131 unit tests, all passing ✅ Real ML models trained (F1=0.7616, ROC-AUC=0.8454) ✅ Production-grade code generation ✅ SHAP feature importance analysis ✅ Automatic CI/CD integration ✅ Completed in 3-4 days

What we learned

• Multi-agent systems require careful state management
• Claude's reasoning is powerful for ML domain knowledge
• GitLab's agent platform enables complex integrations
• SHAP provides excellent model interpretability
• Production ML requires robust error handling

What's next for ML Pipeline Orchestrator Agent

• Support for deep learning (TensorFlow/PyTorch)
• Model monitoring and drift detection
• A/B testing framework
• Automated model registry integration
• Multi-class and regression task support

Built With

• claude-sonnet-4.6
• docker
• dotenv
• gitlab-ci/cd
• gitlab-duo-agent-platform
• pytest
• python
• scikit-learn
• shap
• xgboost