Lesson 10: Capstone: build a small, reviewable AI workflow

Lesson 10 of 10

Capstone: build a small, reviewable AI workflow

Build a prototype that sorts public, non-sensitive feedback into a few topics. Start with a clear goal and a small consented or synthetic dataset. Define categories, create a baseline, divide examples without leakage, and evaluate errors before showing predictions to a reviewer.

1. Scope: State what the prototype can and cannot decide.
2. Data: Record where examples came from and remove unnecessary identifiers.
3. Baseline: Try simple keyword rules before selecting a learned model.
4. Evaluation: Measure per-category precision and recall; inspect mistakes.
5. Review: Present the suggested topic with the original text and allow correction.
6. Document: Note limitations, data handling, model version, and rollback steps.

for feedback in new_feedback:
    suggestion = topic_model.predict(feedback.text)
    show_to_reviewer(
        original=feedback.text,
        suggested_topic=suggestion,
        allow_correction=True
    )

Do not automatically make high-impact decisions from the prototype. A practical AI project is successful when it solves a defined problem reliably, can be checked by people, and has a safe fallback.

Capstone deliverable: Write a one-page project brief with the goal, data source, baseline, evaluation measures, likely failure modes, human review plan, and next experiment.