What You Will Be Able To Do

• Write a one-page project brief.
• Combine prompt, knowledge, and API.
• Evaluate with a fixed test set.
• Deliver a clear five-minute demo.
• Review a peer project.

Recap Of Week 11

• Think of jobs as tasks; checking skill gains value.
• Use AI to learn, not to skip learning.
• People stay responsible for AI decisions.
• The EU AI Act regulates by risk tier.

The Path From Week 1 To Week 12

• 1 to 3: AI, machine learning, neural nets.
• 4 to 6: transformers, LLMs, prompting.
• 7 to 9: knowledge, agents, evaluation, safety.
• 10 to 11: building and society.

concepts -> prompting -> knowledge + tools
   -> evaluation + safety -> API app
   -> responsible use -> CAPSTONE

Small Beats Ambitious

• One user, one problem, one clear task.
• About six to ten hours of work.
• A reliable small tool beats a flaky large one.
• Cut scope early, not late.

Choosing A Project

• Document helper with citations.
• Quiz maker with JSON output.
• Meeting notes to tasks.
• Planner that uses a tool.
• Writing coach with a rubric.

Three Feasibility Questions

• Can you collect 20 realistic inputs?
• Can you tell if an answer is good?
• Can v1 run on a laptop with an API key?
• If any answer is no, shrink the project.

The One-Page Brief (1)

• User: who and in what situation.
• Problem: what is slow or hard today.
• Solution: two sentences.
• Design: prompt, knowledge, model, reasons.

The One-Page Brief (2)

• Success criteria with numbers.
• Top three risks with a guard each.
• Cost and speed targets.
• Out of scope.

Success: >= 16 of 20 test questions correct,
         every answer cites a source.
Cost:    < $0.02 per answer
Speed:   < 5 s to first answer

Build In Thin Slices

• 1: system prompt plus API call, test on 5 inputs.
• 2: add retrieval or one tool.
• 3: add structured output with validation.
• 4: add guards and privacy handling.
• Run the test set after every slice.

Keep A Build Log

• Date, change, score, cost.
• One change at a time.
• Save failing examples.
• The log becomes your report.

date    change              score  cost
Mon     bare prompt         9/20   $0.01
Tue     + retrieval k=3    14/20   $0.02
Wed     + 'not found' rule 16/20   $0.02

Evaluation

• At least 20 cases, fixed before tuning.
• Include 3 that should be declined.
• Score by a rule written in advance.
• Report failures honestly.

The Rubric

• Problem and user; design; working demo.
• Evaluation; safety and privacy.
• Responsible use; code and cost.
• Presentation. Each row 1 to 4, total 32.

Demo Script

• 0:00 user and problem as a story.
• 0:30 what it does, one diagram.
• 1:00 live demo, happy path.
• 2:30 one failure on purpose.
• 3:15 results, risks, next step.

Backup plan: demo input in a text file,
recorded video, screenshots of results.

Rehearse And Prepare

• Rehearse twice with a timer.
• Paste prepared inputs; avoid live typing.
• Hide keys and private data on screen.
• Record a backup video.

Peer Review Checklist

• Clear user and problem within 30 seconds?
• Demo runs on a fresh input?
• Sources cited or 'not found' used?
• Guards survive two reviewer attacks?
• Cost and speed reported with numbers?

Recap And What Comes Next

• You can prompt, ground, build, and evaluate.
• You can design with safety and privacy.
• Keep learning: one paper or changelog a month.
• Share your project with real users.

Questions And Capstone Tasks

• Write and share your one-page brief.
• Build thin slices and keep the log.
• Rehearse the demo and complete one peer review.