24Capstone: Build a Complete AI Solution
Put it all together. Choose one of three capstone briefs, build it end to end with a local or cloud model, measure it, and turn it into a portfolio case study.
What you'll learn
- Scope, build and test a complete AI solution
- Combine prompting, RAG, workflows, agents and guardrails
- Measure results against a baseline
- Present the project as a case study
Choose your brief
Barefoot Café — 24/7 customer assistant
Goal: answer Messenger/website questions about hours, menu, delivery and reservations; take reservation requests; escalate complaints.
- Knowledge base: hours, menu with prices, delivery zones, policies (write 1–2 pages of realistic data)
- RAG-grounded answers with “I don't know” behavior (Lessons 05, 08)
- Reservation requests saved to a Google Sheet; staff notified on Telegram (Lesson 14)
- Complaint detection → human handoff
- Stretch: run fully local with Ollama + Open WebUI or n8n (Track 4)
Freelancer back-office agent
Goal: an agent that manages invoices and client follow-ups with approval.
- Invoice data in a sheet or CSV; tools to list, calculate and draft (Lesson 09)
- MCP server exposing the data (Lesson 11)
- Weekly scheduled run with a summary sent to your phone (Hermes or n8n — Lessons 12, 14)
- Approval gate before any email is sent
- Stretch: receipt extraction to structured JSON from photos (Lesson 17)
Dental clinic lead & reminder pipeline
Goal: capture inquiries from a form, qualify them, draft replies, and send appointment reminders.
- n8n form → AI qualification JSON → sheet → approval → email (Lesson 14)
- Scheduled reminder workflow the day before appointments
- Privacy design: what personal data is stored, where, and who can access it (Lesson 22)
- Stretch: private document Q&A over clinic policies for staff (Lesson 18)
Build process
Write the pilot brief
Use the template from Lesson 20: problem, solution, scope, success metrics, risks, owner.
Map the system
Draw trigger → steps → outcome. Mark AI steps, tools, data sources and human approvals. Excalidraw or paper is fine.
Prepare knowledge & test data
Write the knowledge base and 20 realistic test inputs, including 5 tricky ones (angry, vague, Taglish, off-topic, missing info).
Build the smallest version end to end
Get one input all the way through before adding features.
Add guardrails
Step limits, validation, error alerts, approval gates, logging.
Evaluate
Run all 20 test inputs. Score accuracy and note failures. Fix, re-run.
Measure
Compare to a baseline: response time, minutes per task, error rate.
Document
One-page runbook: what it does, how to operate it, how to fix common issues.
Write the case study
Problem → approach → results (numbers) → screenshots/video → what you'd do next.
Grading rubric (self-assessment)
| Criterion | Excellent looks like |
|---|---|
| Problem clarity | Specific problem with a measured baseline |
| Grounding | Answers come from provided data; system admits unknowns |
| Reliability | Validated outputs, error handling, no duplicate actions |
| Safety | Least privilege, approvals on external actions, privacy considered |
| Evaluation | 20+ test cases with pass rates before and after fixes |
| Results | Clear before/after numbers |
| Communication | Case study a business owner understands in 2 minutes |
You've finished the Academy 🎉
You now understand how AI works, can write production-grade prompts, build agents and workflows, run models on your own hardware, and plan, measure and sell AI solutions. Keep going with the Resource Library, experiment in the AI Lab, and use the Toolkit with real clients.
Key takeaways
- A complete AI solution combines grounding, workflows or agents, guardrails, evaluation and measurement.
- Build the smallest end-to-end version first, then harden it.
- A test set and a baseline turn a demo into a credible case study.
- Your capstone is your first portfolio piece — present it in business language.
Knowledge check
0 / 3Q1What should you build first?
End-to-end first reveals integration problems early.
Q2How many test inputs does the capstone recommend at minimum?
About 20 realistic cases, including tricky ones, gives meaningful confidence.
Q3A great case study leads with…
Owners care about their problem and results.