The ML Project Lifecycle
A machine learning project is not a single training run. It is a loop that begins with a business problem and never truly ends while the model serves traffic.
The main stages
- Scoping turns a vague goal into a measurable target, such as reducing fraud losses by a set percent.
- Data covers collection, labeling, and cleaning so the model has something to learn from.
- Modeling picks features, trains candidates, and tunes them against a metric.
- Deployment ships the model behind an API or batch job.
- Monitoring watches accuracy and inputs so you catch drift before users do.
Why it is a loop
Production almost always reveals problems the lab missed. New data arrives, distributions shift, and labels age. Findings from monitoring feed back into scoping and data, so teams treat the lifecycle as iterative rather than linear. Skipping stages, especially monitoring, is how a once accurate model quietly rots.
Key idea
The lifecycle is a closed loop where monitoring in production feeds the next round of scoping and data work.