FAST-12 Projects: AI Practitioner Certification Prep (AIF-C01)
This is a certification-prep course, so the projects build exam readiness and a work artifact you can reuse. Use fictional data or a sanitized description of a use case at your job (no customer personal information, PHI, or confidential details; your manager approves the use case). Scoring: Not yet / Meets / Strong. Completion needs Meets on every capstone criterion and 80% or higher on two timed practice exams (one resubmission allowed).
Capstone: AWS Generative-AI Use-Case Brief
What you produce: a 2-page brief recommending how a fictional or sanitized organization should build one generative-AI use case on AWS, plus a 5-minute walkthrough in session 5. The brief exercises every exam domain: fundamentals (Domain 1), generative-AI trade-offs (Domain 2), foundation-model application design (Domain 3), responsible AI (Domain 4), and security and governance (Domain 5).
Required contents
| Section | Content | Domain |
|---|---|---|
| Problem and fit | Business problem; why AI (or why not); success metric | 1, 2 |
| Approach | Prompting, RAG (knowledge base), fine-tuning, or agent; why; what you rejected and why | 3 |
| Model choice | Two candidate models in Bedrock; comparison on quality (your evaluation set), latency, and cost per month (show the math) | 2, 3 |
| Evaluation | Evaluation set of 10+ questions with expected answers; automatic or LLM-as-a-judge plus human review; pass threshold | 3 |
| Responsible AI | Top three risks, the responsible-AI dimension of each, and the control (guardrail, human review, disclosure) | 4 |
| Security and governance | Shared responsibility split; IAM roles; encryption (KMS); logging (CloudTrail, invocation logs); data classification; compliance needs (for example, HIPAA BAA via AWS Artifact, G.L. c. 93H / 201 CMR 17.00 for MA personal information) | 5 |
| Rollout | Pilot scope, owner, review date, rollback trigger | 3, 5 |
Rubric
| Criterion | Not yet | Meets | Strong |
|---|---|---|---|
| Fit and approach | No reason given for the approach, or AI used where a rule would do | Approach justified; one alternative rejected with a reason | Two alternatives compared on cost, effort, and quality |
| Model comparison | No numbers, or cost math wrong | Quality and monthly cost compared with correct math | Latency and context-window limits included; sensitivity to volume shown |
| Evaluation | No evaluation set | 10+ questions with expected answers and a pass threshold | Includes "not in source" and injection tests; human review defined |
| Responsible AI | Risks generic or controls missing | Three specific risks, each mapped to a dimension and control | Residual risk and monitoring plan stated |
| Security and governance | Shared responsibility wrong, or no IAM, encryption, or logging | All five covered correctly for the use case | Least-privilege policy sketched; retention and audit cadence set |
| Walkthrough | Over time or unclear | 5 minutes; recommendation first | Answers two questions using exam vocabulary precisely |
Example (excerpt): Charles River Family Dental (fictional) booking assistant
Problem and fit. Front desk spends about 9 hours a week answering hours, location, and insurance-network questions. Success: 60% of those questions answered by the assistant with 95% correct answers on the evaluation set. Booking itself stays in the practice system (a rule-based flow, no AI needed).
Approach. RAG over the office FAQ and accepted-insurance list using a Bedrock knowledge base. Rejected fine-tuning: the facts change monthly and RAG updates with a re-sync; fine-tuning would need retraining and gives no citations.
Model choice. Model A ($0.80/$3.20 per million tokens, sandbox rates) vs Model B ($3/$15). At 200 questions a day, 1,500 input and 150 output tokens each, 30 days: A ≈ $10.08 a month; B ≈ $40.50 a month. Both passed 19 of 20 evaluation questions; choose A.
Responsible AI. (1) Clinical advice to patients (safety): Bedrock guardrail denied topic plus a blocked message pointing to the office phone and 911. (2) Patient details repeated in chat (privacy): sensitive-information filter masks phone and email; the assistant never sees appointment records. (3) Spanish-speaking patients get worse answers (fairness): evaluation set includes 5 Spanish questions; pass threshold applies to each language.
Security and governance. AWS secures the infrastructure and Bedrock service; the practice owns IAM, guardrail configuration, data, and logging. BAA accepted in AWS Artifact before any PHI is in scope; the assistant is designed so no PHI is needed. Encryption with a KMS key; CloudTrail and invocation logging on, 1-year retention; quarterly review by the office manager.
Check the example's math: 200 × 1,500 = 300,000 input tokens a day; × 30 = 9 M; at $0.80 = $7.20. Output 200 × 150 × 30 = 0.9 M; at $3.20 = $2.88. Total $10.08. Model B: 9 × $3 = $27.00 + 0.9 × $15 = $13.50; total $40.50.
Mini-project 1: Service Map Flashcards (sessions 1-4)
What you produce: a one-page service map of the in-scope services you are most likely to confuse, in your own words, plus 20 self-written flashcards (use case → service).
Requirements: at least 15 services across AI services (Comprehend, Lex, Polly, Rekognition, Textract, Transcribe, Translate, Personalize), generative-AI services (Bedrock, Bedrock AgentCore, Nova, SageMaker AI, JumpStart, Amazon Quick), and security and governance services (IAM, KMS, CloudTrail, Config, Macie, Artifact, Inspector). Each card has one sentence on when to use it and one on a look-alike and how to tell them apart.
| Criterion | Not yet | Meets | Strong |
|---|---|---|---|
| Coverage | Under 15 services, or out-of-scope services | 15+ in-scope services across all three groups | All AI services and all listed security services |
| Accuracy | Any wrong use case | All use cases correct | Look-alike distinctions tested with a partner |
| Own words | Copied descriptions | Own wording, one sentence each | Each card tied to a lab or case |
Example card: Front: "Pull totals and line items from scanned supplier invoices." Back: Amazon Textract. Look-alike: Amazon Comprehend analyzes text you already have (sentiment, entities); Textract gets the text and structure out of documents first.
Mini-project 2: Teach-Back on One Task Statement (session 5)
What you produce: a 3-minute explanation of the task statement you scored lowest on, with one original practice question and its explanation.
| Criterion | Not yet | Meets | Strong |
|---|---|---|---|
| Accuracy | Factual error | Correct and matches the exam guide wording | Cites the AWS documentation page used |
| Practice question | Ambiguous or answer cued by length | One clear best answer; plausible distractors; explanation of each wrong option | Scenario-based, fictional, and peer-tested |
| Delivery | Over 3 minutes | 3 minutes; one example | Uses a lab result as the example |
Example (excerpt): "Task 3.1: choosing between prompting, RAG, and fine-tuning. In Lab 3 our knowledge base missed hearing protection because the PPE table was split across chunks. That's a retrieval problem, so the fix is chunking, not fine-tuning..."