Client capstone: from discovery to a live system
The capstone is the job in miniature. The code matters, but the scope, the metric and the story you tell with numbers matter as much.
The brief
Choose a domain you can get realistic data for: insurance claims, bank support, clinic scheduling or retail returns. Treat it as a customer engagement from the first conversation to the final presentation.
What you will build
- A one-page scope with a single success metric and exit conditions
- A deployed system built from the parts of this roadmap
- An eval report and a live demo
- A results write-up with before and after numbers
Requirements
| Requirement | Done when |
|---|---|
| Scope | One metric, agreed before building, with what is out of scope |
| System | Uses retrieval or agents, guardrails and human approval where actions matter |
| Evidence | An eval suite and traces back every claim |
| Deployment | Runs somewhere a reviewer can use it |
| Story | A ten-minute presentation for an executive and a deeper one for an engineer |
What to learn it from
- Topic: discovery and scoping
- Build the support agent in LangGraph, the OpenAI Agents SDK or Google ADK
- Claude Cowork: the full course, for agents business teams use
Milestones
- Interview someone who does the work, even for 30 minutes
- Write the scope and metric
- Build the thinnest end-to-end version
- Add evals, guardrails and approval
- Deploy it
- Measure, write up and present
How it is judged
Could an interviewer poke holes in any claim in your write-up? Every number needs a trace, an eval or a measurement behind it.
Stretch goals
- Run it with a real user for a week
- Add a second domain and extract a reusable template
- Record the presentation
Before you call it done
Worth remembering
- The metric came before the code
- Every claim has evidence
- You can explain it at two levels
Every expert started right here.