The AI/ML Development Lifecycle
From raw data to a monitored production model: the ML pipeline, deployment options, MLOps practices, and the metrics that prove a model works.
A model that scores well in a notebook is not yet a product. This topic follows the full path that turns data into a model your users can trust, and then keeps that model working after launch. It maps directly to Task Statement 1.3 of the AWS Certified AI Practitioner exam, one of the most concept-dense parts of the first domain.
What This Topic Covers
- the stages of an ML pipeline, from data collection and EDA through preprocessing, feature engineering, training, tuning, evaluation, deployment, and monitoring
- where models come from (pre-trained, customized, or trained from scratch) and how you serve them (managed API versus self-hosted, and real-time, serverless, asynchronous, or batch inference)
- which AWS and Amazon SageMaker capability handles each stage, from Ground Truth labeling to Model Monitor drift detection
- MLOps fundamentals: experimentation, repeatable processes, scalable systems, managing technical debt, monitoring, and retraining
- the metrics that measure success, from the confusion matrix and precision/recall to regression errors and business ROI
Why It Matters
The exam rewards you for seeing ML as a lifecycle, not a single training step. Many questions describe an activity ("the model degraded after launch," "we need labeled data," "score every account overnight") and expect you to place it in the pipeline and name the right practice or service. This topic builds that reflex.
It also closes the first domain by connecting every earlier idea to a result you can measure. Data types, learning styles, and model choices all feed into this pipeline, and the metrics at the end decide whether the whole effort was worth it. Get this lifecycle straight and the later domains, which assume you already think in pipelines and metrics, become far easier.