Automate the ML Lifecycle from Data to Deployment

The ML Engineer Agent is an autonomous agent that manages the complete machine learning lifecycle from data analysis and model development through production deployment and monitoring, handling everything from feature engineering to model serving infrastructure. It works in six stages: analyzing the project to identify the ML problem type, data quality, and success metrics; building data preprocessing and feature engineering pipelines with validation, stratified train/validation/test splits, and data-leakage prevention; developing models by selecting algorithms, implementing baselines, and designing experiment tracking with hyperparameter optimization; training multiple model candidates, tuning hyperparameters systematically, and generating interpretability and feature-importance analyses; preparing for production by packaging models with versioning and metadata, building serving APIs and inference pipelines, and implementing drift detection and automated retraining triggers; and finally deploying to staging and production with logging, metrics, A/B testing frameworks, and performance dashboards.

Use this agent when you need the full ML pipeline handled end to end - not just model training in isolation, but the surrounding data validation, experiment tracking, and production serving infrastructure too. Its guidelines require reproducibility (fixed random seeds, version-controlled data and code, experiment logs), proper cross-validation with holdout sets to prevent overfitting, and production concerns (latency, throughput, scalability) considered from the start rather than bolted on after a model works in a notebook. It is not meant for a quick one-off notebook experiment with no production intent - the standard project structure it produces (data/notebooks/src/tests/configs directories) and its model card and deployment config templates are built for models meant to run in production, monitored, and retrained.

### Model serving configuration
class ModelConfig:
    model_path = "models/model_v1.pkl"
    feature_pipeline = "pipelines/features_v1.pkl"
    batch_size = 32
    timeout = 30
    monitoring_enabled = True
    drift_threshold = 0.1

Given an ML problem, the agent produces a standard project directory structure (data/raw, processed, features; notebooks; src/data, features, models, deployment; tests; configs), a Model Card documenting problem type, primary and validation metrics, features, training data size and date range, and deployment latency/throughput targets, and a deployment configuration like the one above specifying model path, batch size, timeout, and drift threshold.

ML engineers and data scientists who need a model to go from data to production with reproducibility and monitoring built in, not just a trained model handed off with no serving plan. It suits teams that want automatic handling of common ML challenges like class imbalance, feature scaling, and model interpretability, and that need documented model assumptions and limitations alongside proper authentication and data-privacy measures in the serving layer.