3. ML Model Development (26%)

Model development is the phase most candidates feel comfortable with — and that's exactly where the exam exploits overconfidence. At 26%, this domain tests not just whether you understand algorithms, but whether you can select the right approach for a given scenario, configure training efficiently on SageMaker, and evaluate results using the correct metrics. The three sections follow the natural arc: choose an approach (3.1), train and refine it (3.2), then analyze whether it actually works (3.3).

Pay particular attention to the spectrum of modeling options on AWS — from no-code AI services (Rekognition, Comprehend) through managed foundation models (Bedrock, JumpStart) to SageMaker built-in algorithms (XGBoost, Linear Learner) to fully custom training with BYOC containers. The exam presents business requirements and expects you to pick the right point on that spectrum. More complex is not always better — the simplest approach that meets the requirements is almost always the correct answer.