AI-300 study plan

Map the end-to-end model lifecycle (data prep, training, registration, deployment, monitoring) onto concrete workspace artifacts and name the control point at each stage.

Provision a workspace with a compute cluster and register a trained model into the registry with version, tags, and lineage metadata.

Build a multi-step pipeline using reusable components and rerun it with a changed hyperparameter to prove run reproducibility.

Wire a CI workflow that triggers the pipeline on merge using a service-principal connection.

Deploy a registered model to a managed online endpoint as a blue deployment and validate with scored requests.

Implement a 10/90 canary traffic split and shift traffic only after a latency and error-rate check passes.

Configure data drift and performance monitoring with alert rules routed to an action group.

Define and automate a retraining trigger based on drift thresholds or a schedule.

Author infrastructure-as-code that provisions workspace, storage, key vault, and container registry with private endpoints and managed identity.

Assign least-privilege roles for a data scientist persona, a CI service principal, and a scoring endpoint identity.

Estimate monthly cost for the compute footprint and apply quota and budget controls.

Tear down and redeploy the whole environment from code alone to prove repeatability.

Provision an AI hub and project with connected search, storage, and secret-store resources.

Deploy one chat-class model and one embedding model, then compare pay-as-you-go against provisioned throughput on cost and latency.

Build a retrieval-augmented generation flow over indexed documents and publish it behind an endpoint.

Version prompts and flow definitions in Git and require an evaluation run before any promotion.

Implement CI/CD for generative AI assets so evaluation runs gate promotion to production.

Configure content filters, a routing layer across two deployments, and retry/fallback behavior on failure.

Centralize connection and secret management so no credentials appear in code or config files.

Document and rehearse a rollback by reverting to a prior known-good flow version.

Assemble an evaluation dataset of 30+ labeled pairs that includes adversarial, out-of-scope, and ungrounded-answer cases.

Run AI-assisted scoring for groundedness, relevance, coherence, and fluency and record baseline values.

Instrument requests to capture prompt, response, token counts, latency, and tool calls in an observability backend.

Configure continuous evaluation sampling with regression thresholds that fail the build on degradation.

Execute a red-team pass and log safety and jailbreak findings with severity ratings.

Cut token consumption per request by tuning chunk size, enabling hybrid retrieval with semantic ranking, and pruning context.

Benchmark p50 and p95 latency and quality across at least three model variants and select the cheapest that clears the quality bar.

Add response caching and prompt compression, then measure the resulting cost and latency delta.

Write a reasoned RAG-versus-fine-tuning recommendation for a concrete scenario with expected tradeoffs.

Re-run every lab that failed or was completed with a workaround and close the gap without notes.

Take two full-length practice assessments and tag every miss by domain and root cause.

Produce a one-page decision summary per domain covering when to choose each deployment, monitoring, and evaluation option.

Teach back the three weakest subtopics aloud or in writing to confirm retention.