Machine Learning Lifecycle
A machine-learning system is a lifecycle, not a single training job.
Core lifecycle
Problem → data → preparation → training → evaluation → deployment → monitoring → feedback → iteration
1. Frame the problem
Define the business outcome, prediction target, constraints, baseline, and what happens when the model is wrong.
2. Build the data pipeline
Collect representative data, define labels where required, validate quality, and prevent leakage between training and evaluation datasets.
3. Train and evaluate
Train against a reproducible dataset and compare against a simple baseline. Choose metrics that reflect the real failure costs rather than relying on accuracy alone.
4. Deploy
Choose batch, asynchronous, online, or edge inference based on latency, volume, freshness, and operational requirements.
5. Monitor
Track service health and model behavior: latency, errors, input drift, output quality, data quality, and business KPIs.
Production checklist
- Reproducible training inputs and code
- Versioned model artifacts
- Explicit evaluation dataset
- Rollback strategy
- Monitoring and alert thresholds
- Ownership for retraining and incident response
Connected knowledge
datasets? → feature engineering → LLMs → RAG → production agents.