The Capstone Brief
Choose a real community use case and plan an architecture package judged first on judgment: when AI is used, when it is not, and who is protected.
Objective
By the end of this lesson, you can scope a capstone use case with a community organization, and explain how your architecture package will be judged.
What you will hand in
Your capstone is not a demo. It is the package a careful organization needs before anyone writes production code:
- An architecture document that explains and justifies every component.
- A risk register that names each risk, scores it, and gives it an owner and a mitigation.
- An evaluation plan with test sets, thresholds, and launch and stop criteria.
A working prototype is welcome, not required. A clear "do not build the AI part" decision, backed by evidence, can earn full marks.
The default use case: a nonprofit intake assistant
Many community organizations spend staff time answering the same questions and collecting the same information at intake. Across the U.S., people also reach local services through 211, the helpline run by United Way and partner agencies. Imagine a neighborhood nonprofit that offers food, rental help referrals, and tax help. It wants an assistant on its website and text line that:
- Answers common questions about services, hours, and documents to bring.
- Collects basic intake details and routes the request to the right staff member.
- Works in English and Spanish, on low-end phones and slow connections.
- Never decides eligibility, and always offers a person.
You may choose a different organization: a school family center, a library, a clinic front desk, a youth program, a tenant association. Get the organization's permission if you name it. Use only made-up or public data. Never use real client records, and never use data about real minors.
Capstone
Capstone brief: an AI architecture that earns trust
- A community organization: a neighborhood nonprofit by default, or one you choose with its permission
- Design an AI-assisted intake and information system that helps people get the right service faster, protects their privacy, and never leaves anyone worse off for not using it.
Deliverables
- Architecture document (6 to 12 pages): problem statement, non-AI baseline, build/buy/blend ADR, component diagram, data flow map, prompt and tool contracts, RAG design, agent and approval design, security and privacy design, reliability and fallback design, deployment and monitoring plan
- Risk register: at least 12 risks with likelihood, impact, score, owner, mitigation, and the signal that tells you it is happening
- Evaluation plan: test set design with case types and languages, metrics, launch bar, never-regress list, judge calibration plan, red-team plan, and stop criteria
- A 10-minute presentation to a panel that includes a community member, with time for hard questions
| Criterion | Meets the standard when |
|---|---|
| Judgment (weighted most) | Says clearly where AI is not used and why; keeps people in control of every consequential decision; names who could be harmed and how the design protects them, including minors |
| Equity and access | Full non-AI path; works in every language served and on low-end phones; fairness metrics split by group |
| Privacy and security | Data flow map with consent and retention at each stop; least privilege; injection and exfiltration addressed; no secrets in prompts |
| Evaluation | Test set built before features; thresholds and stop criteria set in advance; LLM judge calibrated against people or not used |
| Technical soundness | Each component is justified against a simpler option; contracts, retrieval, fallbacks, and monitoring are specific enough to build from |
| Honesty and clarity | No invented numbers; assumptions labeled; plain language a board member can follow |
Judgment comes first in the rubric on purpose. A simple design that knows its limits beats an impressive design that ignores who gets hurt.
Scope it this week
Scope your capstone
0 of 7 doneWho could be hurt?
Go deeper: AI, Minus the Hype: A Practical Guide to Solving Real Business Problems
Discussion
Ask. Answer. Build together.
Loading…