AI curriculum
Course outline for Cloud Foundations for AI Engineers
Explain how a small booking application uses compute, storage, networks, and identity. Compare infrastructure, platform, and software services by what the customer controls and operates. Shortlist ways to run one AI feature, compare their quiet and busy costs, then justify a choice. Restore actual local application data and compare the result with recovery targets set before the drill.
Cloud Foundations for AI Engineers Course Objectives
- Map application requests across compute, storage, networking, and identity boundaries.
- Distinguish IaaS, PaaS, and SaaS through customer control and retained operating duties.
- Compare managed and serverless services using lifecycle, persistence, limits, and responsibility.
- Choose hosted inference, a managed AI platform, or virtual machines for the same workload.
- Distinguish inference, fine-tuning, and foundation-model pretraining without assuming universal platform support.
- Calculate two-utilization total costs with sourced rates, ancillary charges, and staffing assumptions.
- Execute a local backup/restore drill and verify restored records independently.
- Compare observed recovery time and lost changes with stated recovery targets.
Pre-requisites
- Describe a simple browser request and distinguish application code from stored data.
- Use files, a browser, a terminal, and basic arithmetic.
- No cloud account, AI terminology, or provider experience is required.
Before this course
Complete the required courses or bring equivalent experience. Recommended courses add useful depth; they are not enrolment gates.
No prerequisite course is required. The entry skills above still apply.
Readiness checks
- Describe a simple browser request and distinguish application code from stored data.
- Use files, a browser, a terminal, and basic arithmetic.
- No cloud account, AI terminology, or provider experience is required.
Describe a simple browser request and distinguish application code from stored data.
Find your place in the learning mapLab Setup
Delivery requirements: the environment, access, and exercise materials must be prepared and verified before delivery. This outline does not certify that a lab package is ready.
Local booking application and supported runtime behind a generic backup/restore interface; synthetic records; separate disposable restore paths; invalid-backup case; editable worksheets; dated generic capability, responsibility, price, and data-handling profiles. Stable runtime support, consistent backups, record checks, complete pricing evidence, formulas, fault trials, and cleanup require verification. No vendor-specific SDK or named database track is required.
Detailed Course Outline
Module 1: Follow one ordinary application request (2 hours)
- Compute lifecycle: Processes, virtual machines, compute capacity, and resource lifecycle.
- Storage roles: Object, file, block, and database storage roles.
- Request paths: Addresses, names, routes, and ports along one request.
- Reachability and authorization: Public/private boundaries and network reachability versus authorization.
- Identity and least privilege: Users, service identities, credentials, roles, and least privilege.
- Resource ownership: Browser-to-application-to-data model with explicit resource owners.
- Guided-time allocation: 0.75 hours for explanation and a worked example, 0.75 hours for learner practice, 0.25 hours for feedback, and 0.25 hours for questions and contingency.
Practical outcome: Draw and explain a booking request and its separate network, identity, and persistence boundaries.
Module 2: Decide what the provider operates (2 hours)
- Cloud characteristics: Cloud characteristics and service models versus public/private deployment models.
- Infrastructure as a Service: Infrastructure as a Service: guest systems, applications, and patching ownership.
- Platform as a Service: Platform as a Service: deployed applications and retained configuration responsibilities.
- Software as a Service: Software as a Service: provider applications and customer access/data decisions.
- Managed and serverless state: Managed databases, serverless limits, temporary runtime data, and records that must survive.
- Retained operating duties: Operating duties, recovery needs, and data location for the same service choice.
- Guided-time allocation: 0.75 hours for explanation and a worked example, 0.75 hours for learner practice, 0.25 hours for feedback, and 0.25 hours for questions and contingency.
Practical outcome: Compare IaaS, PaaS, and SaaS and explain who retains each operating and data responsibility.
Module 3: Shortlist where the booking-help model runs (2 hours)
- Fixed AI workload: One booking-help workload with fixed quality, latency, volume, and data restrictions.
- Three consumption routes: Hosted ready inference API, managed AI platform, and raw-VM self-hosting.
- Inference fine-tuning and pretraining: Inference requests, fine-tuning existing weights, and foundation-model pretraining.
- Managed platform capabilities: Provider model access and supported changes versus broader model-development workflows.
- Model and resource restrictions: Model licenses, GPU memory, endpoint capacity, quotas, scaling, and availability.
- Provisional shortlist: Operating duties, data restrictions, and a provisional feasible shortlist.
- Guided-time allocation: 0.75 hours for explanation and a worked example, 0.75 hours for learner practice, 0.25 hours for feedback, and 0.25 hours for questions and contingency.
Practical outcome: Classify new model-operation needs and produce a provisional feasible shortlist for the same booking-help workload.
Module 4: Calculate quiet and busy operating cost (2 hours)
- Workload and utilization: Request volume, simultaneous requests, duration, capacity, and used versus available capacity.
- Fixed and request charges: Fixed instance hours versus request and execution-duration charges.
- Ancillary costs: Storage, requests, transfer, logs, backups, and operating effort.
- Dated pricing assumptions: Region, currency, pricing date, billing tiers, and free-tier assumptions.
- Three-route cost comparison: Quiet and busy cost worksheet for ordinary compute and three AI routes.
- Budgets and simpler alternatives: Budget alerts versus enforced limits and a simpler local alternative.
- Guided-time allocation: 0.75 hours for explanation and a worked example, 0.75 hours for learner practice, 0.25 hours for feedback, and 0.25 hours for questions and contingency.
Practical outcome: Compare ordinary compute and all three AI routes at quiet/busy levels, then justify the final choice.
Module 5: Restore data before trusting a backup (2 hours)
- Availability and recovery: Availability, durability, backup, replication, and recovery as different concerns.
- Advance recovery targets: Recovery-time and acceptable data-loss targets written before the drill.
- Consistent backup checkpoint: Local backup with the last saved booking included as its checkpoint.
- Separate restore location: Disposable loss/corruption injection and actual restore into a separate location.
- Independent recovery checks: Independent record comparison, elapsed time, and missing-change observations.
- Failed restore and retry: Failed restore diagnosis, clean retry, and cloud-recovery evidence gaps.
- Guided-time allocation: 0.75 hours for explanation and a worked example, 0.75 hours for learner practice, 0.25 hours for feedback, and 0.25 hours for questions and contingency.
Practical outcome: Restore local records and compare measured recovery with advance targets, while identifying untested cloud risks.
Practical exercise
- Use an ordinary booking application with a prepared local runtime and database data. Planned starter materials include the app, sample records, backup/restore hooks, an editable worksheet shell, and source-backed service-profile cards. Learners author the resource/request model, responsibility choices, formulas, recovery procedure, and one independent data check. These materials must be prepared and run before delivery; they are not claimed as delivered assets.
- Map successful booking requests and predict separately what happens when a route fails or the identity lacks permission. These network/identity cases are diagrams and reasoned checks, not deployed cloud policy tests. Compare operating booking code on a VM, deploying it on a managed runtime, and consuming a complete hosted booking product. Add managed database/storage and serverless handler choices. Do not put the only booking copy in temporary function storage. Explain provider and customer duties rather than assuming managed means responsibility-free.
- For the booking-help feature, compare three alternatives against the same request set, quality floor, response-time requirement, data classification, quiet volume, busy volume, and service hours. Use a ready inference API, a managed AI platform, and a raw-VM self-hosted runtime. No vendor account is required. A complete SaaS application is a separate option if it actually meets the need. Do not classify an API as SaaS only because it has an API. State whether the task needs inference alone or a justified later customization. Hold training out of the base inference estimate and itemize any proposed training and evaluation work separately. Reject pretraining when no requirement justifies it.
- Complete profile fields for supported operation, model/license restrictions, endpoint and geographic processing path, data retention and training-use terms, access permissions, quotas, scaling/cold-start behavior, idle charges, observability, maintenance ownership, retained storage, recovery, staffing, and unresolved facts. Separate provider documentation from instructor workload assumptions. Unknown throughput, account eligibility, or retention requirements remain unknown. An unresolved hard data restriction blocks a recommendation until evidence resolves it. Do not assume the same quality, model weights, performance, or privacy across routes.
- Build an editable quiet/busy worksheet for ordinary always-on versus request-driven compute first. Record request rate, duration, memory, capacity, billing interval, region, currency, dated price source, and units. Include storage, operations, transfer/egress paths, logs, backups, and operating effort. Preserve this basic cloud comparison before extending the worksheet to the three AI consumption routes. Record input/output token units for APIs, endpoint or VM hours for provisioned capacity, minimum billing/idle costs, any scaling assumptions, and ancillary charges. Include model/runtime licensing fees where applicable, setup and maintenance effort, on-call staffing, and support assumptions. Staff rates are declared assumptions, not vendor quotes. Avoid counting the same storage or backup twice.
- Separate cash charges from estimated staffing cost. Show both total cost and cost per successful booking or accepted help answer with explicit denominators. Do not compare a per-token rate directly with a VM-hour rate. Vary volume, duration, storage, operator effort, and utilization. Check one formula independently. If quality or capacity evidence is missing, label conditional estimates rather than inventing completed-answer throughput. No route is universally cheapest. List prices precede account-specific credits/free tiers. Alerts are not immediate hard spending caps. Actual bill measurement and model benchmarking are not claimed.
- Before the drill, write the maximum acceptable recovery time and data-loss window. Fix the clock start at fault injection and the end at successful application restart plus independent record checks. Include diagnosis and retries in that interval. Record saved-change identifiers, values, and times for the loss comparison. Do not reset the targets or clock after seeing a failed attempt. If verification never completes, report incomplete recovery rather than a successful duration.
- Use the same advance targets and timing rules for two separate fault trials. The first restores the valid backup. The second starts a new disposable fault, attempts the invalid backup, and retries with the good copy. Record both complete intervals separately. Do not restart the second trial's clock at the safe retry.
- Create a real local backup using a consistent database backup procedure. Record the last included booking, make a later change, then damage only the disposable working copy. Restore to a separate path, restart the app against it, and check record identities and values against the expected checkpoint. Measure elapsed recovery time and identify lost later changes. Attempt an invalid backup, observe failure without destroying the good copy, and recover from the valid one. Compare the saved changes absent after restoration with the predeclared loss window. Use identities and values, not counts alone. Conclude met, not met, or unknown for each target, with the observations that support the conclusion. Missing timestamps or an incomplete check leave the corresponding conclusion unknown. Predict one new failure and write its check before opening feedback.
- For the independent final recommendation, evaluate a changed volume, data restriction, or staffing limit after Module 4's cost comparison. Learners submit their Module 3 shortlist, all three quiet/busy estimates, the final choice, rejected option, source evidence, and unknowns before feedback. A changed requirement needs the affected estimates and eligibility checks updated. Do not omit a route merely because the provisional shortlist rejected it. Discuss provider outage and quota exhaustion without assuming self-hosting removes hardware or regional failures. Define bounded waiting, manual service, or refusal when no permitted route works. Do not propose a fallback that violates the same data or quality requirements.
- The restore demonstrates local restoration only. It does not test cloud account loss, regional outage, remote backup reachability, compromised credentials, provider logs, GPU replacement, production scale, or full disaster recovery. Keep training datasets, model files, and required configuration in a proposed recovery inventory only when the selected consumption route actually owns them. Provider redundancy is not proof that customer application data can be restored. Close by linking the selected route's recovery inventory to the drill: state what the local check proves and which remote/provider dependencies remain untested. Do not claim the local result validates the selected cloud architecture.
Delivery and timing
- Guided hours include teaching, demonstrations, learner practice, feedback, questions, and contingency as allocated above. Breaks, bulk downloads, account approval, and prerequisite setup are additional. Allocations need a timed learner pilot before delivery; they are not measured completion times.
What you will produce
Compare cloud responsibilities and AI consumption choices, estimate workload costs, and demonstrate bounded local recovery.
How the work is checked
- The ordinary application model separates compute, persistence, routes, and identities.
- IaaS, PaaS, and SaaS comparisons use control and responsibility, not interface labels.
- Managed/serverless choices name retained duties, duration limits, and ephemeral-state limits.
- A provisional shortlist preserves all three routes and marks blocked or unresolved requirements.
- All three quiet/busy cost estimates inform the independent final justified choice.
- New inference, fine-tuning, and pretraining classifications precede explanatory feedback.
- Proposed model changes include supported-method and data-requirement checks.
- Managed-platform comparisons distinguish access, supported customization, training, and serving.
- Licensing, GPU memory, quotas, scaling, data handling, and staffing affect the recommendation.
- Quiet/busy formulas retain units, sourced rates, ancillary charges, and explicit cost exclusions.
- Per-token prices do not masquerade as comparable total operating costs.
- The learner performs a local backup and actual restore, not a trace review.
- Advance targets and a fixed fault-to-verified-use interval prevent target selection after the result.
- Record identities, values, and lost-change times support met/not-met/unknown recovery conclusions.
- A new service choice and recovery check precede explanatory feedback.
- Failures, safe retry/manual alternatives, and untested cloud risks remain explicit.
10 guided hours. General cloud foundations precede AI transfer. Service choices stay vendor-neutral. An API interface does not by itself establish SaaS. Inference, supported customization, and foundation pretraining are different tasks. Service profiles support decision practice, not deployed cloud policy tests or measured model performance. Actual local backup, restoration, restart, and independent record checks are required. They do not certify cloud disaster recovery. No provider account, model call, GPU, or cloud provisioning is required.
When another approach is better
- Keep a simple local deployment when cloud operating cost and complexity exceed the workload's needs.
- Do not choose an ephemeral runtime for the only copy of essential application data.
How we train
Contact us for full course details, including duration, delivery options and lab requirements.
Ask about exercises, instructor feedback and the prior knowledge you need. Tool-specific courses marked provisional may become modules in a broader course.
Discuss your learning goalsRelated Courses
These courses list this course as preparation.