Your Capstone Brief
The brief, the deliverables, and the rubric for your final project, with judgment scored first.
Objective
By the end of this lesson, you can choose a capstone question, plan your project using the seven steps from Module 1, and explain how it will be judged.
The project
You will build one complete machine learning or AI project on public data, for a question that matters to a real community or organization. You will show every step: question, data, preparation, exploration, model, test, and review.
The goal is not the most advanced model. The goal is a model that earns trust: honest about its error, fair across groups, clear about its limits, and stopped if it should be.
Capstone
Capstone brief: a model that earns trust
- A community organization, public agency, or youth program you choose (a real one or a realistic composite; no private data from it)
- Pick a question this organization faces that public data can inform, such as forecasting demand for a service, finding places with low internet access, or sorting public notices. Build, test, and review a model that helps answer it, and recommend whether it should be used.
Deliverables
- A one-page project brief: the question, who decides, who is affected, and the public dataset(s) with sources and licenses
- A Google Colab notebook (or Python files) that runs from start to finish: cleaning log, charts, baseline, model, test results
- An evaluation table: overall error, error by at least one group, and the five worst misses with explanations
- A judgment plan: the five-question test from Module 13, privacy steps, and a build, change, or stop recommendation
- A model card: one page on what the model is for, how it was tested, where it performs worse, and what it must not be used for
- A 7-minute presentation with captions or a written script, for the review board
| Criterion | Meets the standard when |
|---|---|
| Judgment | Names who is affected, applies the five-question test honestly, and makes a clear build, change, or stop recommendation backed by evidence, including when not to use the model |
| Data integrity | Public, licensed data with sources cited; every cleaning choice logged; no personal data and nothing about minors |
| Honest evaluation | Beats a stated baseline on a held-out test set; reports error in plain units, by group, and its worst misses; no data leakage |
| Technical execution | The notebook runs from top to bottom without errors, and the model choice fits the problem |
| Clarity and accessibility | Plain language a community member can follow; charts are honest, labeled, and have text summaries; captions or script provided |
Work alone or in a pair. Pairs submit one project and both present. If you are under 18, a parent or guardian should know your topic and data source.
Choosing a good question
Good capstone questions:
- Use place-level or public data (counties, days, stations, public notices), not data about individual people.
- Help a person make a decision. They do not make the decision for them.
- Have a clear way to measure being wrong.
Example directions, using datasets from this course:
| Direction | Public data |
|---|---|
| Forecast daily electricity demand for one region during heat waves | EIA Hourly Electric Grid Monitor, NOAA daily weather |
| Find counties where internet access lags what income would predict | Census American Community Survey |
| Group counties by food access to help plan mobile pantry routes | USDA Food Environment Atlas |
| Summarize public weather alerts in plain language, with a test set | National Weather Service public alerts |
| Predict heat days for a county to help plan cooling centers | NOAA GHCN-Daily |
Questions to avoid: predicting anything about a specific person's behavior, risk, or worth; anything using data about minors; anything where you cannot measure error.
Proposal check (due at the first capstone Cook Up)
0 of 6 doneReflection
Go deeper: Machine Learning, Minus the Hype: A Practical Playbook to Ship Useful Models
Discussion
Ask. Answer. Build together.
Loading…