Follow the work of building, testing, releasing and maintaining an AI system.
About This Course
A model needs a repeatable path into operation
Getting a model to work in an experiment is an achievement. Keeping a service working as its data, users, software, and business conditions change is another job. Machine-learning operations, or MLOps, connects learning with repeatable releases, monitoring, and recovery. It describes the responsibilities and practices involved in operating the system.
We will build a Production ML Lifecycle Map around two examples: a retailer's fraud model and a document research assistant. The fraud model supports review decisions with predictions. The assistant combines retrieval, generation, and application logic. Both need versions, clear ownership, evidence about performance, and a way to learn from what happens in use.
Review How Machine Learning Learns for training and evaluation and Accelerator Clusters And The Cloud/Software Bridge for serving and observability. You can draw the complete workflow on paper before writing code. The important question is where behavior changed and what evidence would help you choose a response.
Planned Lessons
The Lifecycle Begins With A Useful TaskChoose the problem an AI system should solve and how success will be checked.
Data Pipelines Are Controlled TransformationsFollow the steps that turn raw records into information a model can use.
Features And Feature Stores Need Time AwarenessSee why a prediction must use only information that was available at the time.
Experiment Tracking Records What Produced The ResultKeep a record of the examples, settings and code behind a model’s results.
Reproducibility Has Several Useful LevelsUnderstand what another person needs to repeat an AI experiment.
A Registry Records Candidates And Release StateTrack model versions and the checks that support choosing one for release.
Evaluation Gates Connect Evidence To ReleaseDecide which tests a model must pass before people can use it.
CI, CD, And Continuous Training Are Different MechanismsCompare updating application software, releasing changes and teaching a model again.
Shadow, Canary, And A/B Releases Answer Different QuestionsCompare ways to test a new version safely before sending it to everyone.
Monitoring Covers Service And OutcomeWatch both the health of an AI service and the results people get from it.
Data Drift And Concept Drift Require DiagnosisFind out why a model’s results may change as the world and its inputs change.
Feedback And Labels Reflect Collection ChoicesSee how collecting feedback shapes what a model can learn next.
LLMOps Versions The Evidence And Behavior BundleKeep track of the model, prompts and supporting information behind an AI answer.
Agent Operations Adds State And External EffectsRecord what an AI agent remembers, which tools it uses and what it changes.
Diagnose An Incident Across The ChainFollow an AI problem across its data, model, application and computing services.
Choose Tools By Responsibility And IntegrationChoose tools around the work a team actually needs to manage.