Module 6 · The Model Lifecycle: Train, Evaluate, Deploy, Monitor
The End-to-End Lifecycle
70 min
Learning objectives
- Describe the stages of the ML lifecycle from problem framing to monitoring and iteration
- Explain why the lifecycle is a loop, not a one-time project
- Identify where the most common project failures actually occur
A model is a loop, not a launch
Many people picture building an AI model as a one-time event: gather data, train, ship. In reality, a model lives inside a cycle that keeps turning. The world changes, data shifts, and the model needs care long after launch. Understanding the whole loop is what lets a practitioner judge whether a project is healthy or quietly failing.
Analogy
Think of a model like a garden, not a sculpture. A sculpture is finished once carved. A garden needs ongoing watering, weeding, and replanting as the seasons change — and it dies if you walk away after planting day.
The six stages
- Problem framing — define the decision the model should support and what 'good' looks like.
- Data — collect, clean, label, and split data into training and test sets.
- Train — fit a model so it learns patterns from the training data.
- Evaluate — measure performance on held-out data the model never saw during training.
- Deploy — put the model into a real system where it makes predictions on live inputs.
- Monitor & iterate — watch performance over time and retrain when it degrades.
The arrow from 'monitor' loops back to 'data' and 'train.' That feedback loop is the whole point — a deployed model generates new data and new lessons.
Problem framing — Defining the real-world decision a model supports, the target it predicts, and the measurable definition of success — before any data is gathered.
Problem framing is the stage people skip and regret. If you cannot state what decision the model improves and how you will measure success, no amount of fancy modeling will save the project. A model that predicts something accurate but useless is a failure.
Example — Framing gone wrong
A team built a highly accurate model to predict which support tickets were 'urgent.' It shipped and nobody used it — because urgent tickets were already obvious to agents, and the real bottleneck was routing tickets to the right specialist. Accurate model, wrong problem.
Watch out
Data work — collecting, cleaning, and labeling — typically consumes the majority of a real project's effort. Teams that assume the model is the hard part are usually surprised. Garbage data produces a garbage model no matter how good the algorithm.
Always evaluate on data the model did not train on. Grading a model on its training data is like giving students the exam answers in advance — the score is meaningless.
Knowledge check
Quick practice — not part of your exam score.
Why is the ML lifecycle best described as a loop rather than a straight line?
A team builds an accurate model, but after launch no one uses it because it solves a problem users didn't actually have. Which lifecycle stage failed?
Why must a model be evaluated on data it did not see during training?
Sign in to track your progress and mark lessons complete.
Sign in