Capstone: From Map to Monitored Automation

What you will learn

Choose a capstone process that fits the time budget and shows every part of the method.
Produce the seven required deliverables to the standard in the rubric.
Review a peer's capstone with the checklist and give useful feedback.
Present a workflow, its controls and its numbers in five minutes.

Prerequisites: All of Module 1 to Module 12. The Module 1 task inventory, the Module 4 future-state map and the Module 10 gated flow are the natural starting points.

1. The brief

Build, alone or in a pair, a working automation for one real process from your own work or a provided scenario. It must connect at least three tools, use at least one AI step, gate every action of Tier 3 or above behind a person, log what it does, and come with evidence of how accurate it is. Then present it in five minutes to the group.

The capstone is not judged on ambition. A small process done completely, with every control in place and real numbers, scores higher than a large process with missing gates. Your marker is asking one question throughout: would I let this run in my organisation on Monday?

Time budget

Session 13 is two hours of supervised build time. Most learners also need two to three hours of their own time between Sessions 12 and 13. Choose scope accordingly: if the map takes more than 30 minutes, the process is too big.

2. Marking rubric

Component	Weight	Full marks look like
Workflow map	20%	Current-state and future-state maps with swimlanes, every hand-off shown, each step tagged rule-based / AI / human, and a one-line reason for each automation choice.
Build	30%	A working flow with three or more connected tools and at least one AI step that classifies, extracts, summarises or drafts. Runs end to end on a live test during the presentation.
Controls	25%	Every tool tiered; an approval gate on all Tier 3+ actions with the full action visible to the approver; safe timeout; least-privilege credential; append-only audit log with approve, reject and timeout rows.
Monitoring and evaluation	15%	AI log in place; a golden set of at least 15 items with an accuracy score; a failure branch to a human queue; a kill-switch instruction.
Presentation	10%	Five minutes: the problem, the map, a live run, the controls, the numbers, what you would do next. Answers two questions clearly.

Pass mark for the capstone is 60%. A capstone with no approval gate on an external action caps at 50% regardless of other marks, because it would not be safe to deploy.

3. Required deliverables

Submit a single folder named capstone-<your name> containing:

Maps. Current-state and future-state diagrams (PNG or PDF) with the step-tagging table.
Workflow export. The n8n JSON (or equivalent) plus a screenshot of the canvas with notes on tiers.
Approval evidence. A screenshot of an approval request showing the full proposed action, and the audit log with at least one approved, one rejected and one timed-out row.
Golden set and score. The sheet with agreed labels and the accuracy calculation, plus the changelog showing at least one scored change.
Dashboard. The one-sheet dashboard with runs, failures, cost and the kill-switch instruction.
Risk assessment. One page: the process, data classes involved, tiers, gates, what could go wrong and how it is caught, who owns the workflow.
Slides. Five slides at most for the five-minute presentation.

4. Project menu

Use your own Module 1 task if it fits. Otherwise pick one of these; each has been sized for the time budget.

Leave-request handling

Form in, AI extracts dates and reason and checks the policy text, manager approval gate, calendar entry and confirmation email out. Tier 5 decision stays with the manager.

Vendor onboarding

New-vendor form, AI checks the submitted documents for completeness and summarises them, procurement approval, record created in the vendor sheet, welcome email drafted for approval.

Meeting notes to actions

Transcript or notes uploaded, AI extracts action items with owners and dates, owner reviews the list (edit-then-approve), tasks created in the tracker, summary posted to the channel.

Invoice intake

Invoice PDF arrives by email, AI extracts supplier, amount, PO number and due date, mismatch with the PO sheet escalates to a person, matched invoices are queued for payment approval.

Ticket triage

The Module 6 flow completed: classify, route, draft, approve, log, measure. A good choice if you want depth over novelty.

Weekly status report

Scheduled trigger, pull updates from three trackers, AI drafts the report in the team template, owner edits and approves, report sent and archived.

5. Peer-review checklist

You will review one other capstone before the presentations. Use exactly this list, mark each item yes or no, and write one sentence of the most useful improvement.

[ ] Can I understand the process from the future-state map alone, without explanation? [ ] Is every step tagged rule-based / AI / human, with a reason? [ ] Does the flow run end to end on a fresh test input? [ ] Is every tool tiered, and does every Tier 3+ action have a gate? [ ] Does the approval message show the full action (recipient, text, amount)? [ ] What happens on timeout? Is it safe? [ ] Is the credential a dedicated one with limited scope? [ ] Does the audit log contain approve, reject and timeout rows? [ ] Is the golden set representative, and was it labelled by more than one person? [ ] Is there a failure branch to a human queue? [ ] Is the kill switch written down with a named person? [ ] Does the risk assessment name the data classes and where the model runs? Most useful improvement: ______________________________________

6. Common mistakes that cost marks

Too big a process. Eight swimlanes and twelve tools, none finished. Cut to the three-tool core and say what the next phase would add.
The gate that shows nothing. "Approve? Yes / No" with no content. The approver must see the exact action.
Personal credentials. The flow runs as you, with access to everything. Create the service account; it takes ten minutes.
Golden set from the prompt. The examples in the instructions are the same as the test set, so accuracy is meaningless.
No timeout path. The approval request waits forever, or worse, sends on silence.
An agent where a workflow would do. Autonomy added for show, with no iteration limit. Use the least autonomous design that works, as in Module 1.
Presenting the tool instead of the process. Nobody remembers the node names. Lead with the problem, the numbers and the controls.

Practical lab

The lab is the capstone itself. The steps below are the recommended order with a time budget that fits the two-hour session plus your own preparation time.

1

Choose and scope (20 min, before the session)

Pick the process. Write the one-line goal, the trigger, the final output, and the three tools. If you cannot write those four things in five minutes, choose a smaller process.

2

Map (30 min, before the session)

Draw the current state with swimlanes. Draw the future state. Tag each step rule-based, AI or human and give the reason. Tier every tool.

3

Build the happy path (40 min)

Trigger, connectors, the AI step, the output. Test with one real input until it runs end to end. Use a dedicated credential from the start.

4

Add the gate and the log (25 min)

Approval on every Tier 3+ action, showing the full content. Safe timeout. Audit log with three kinds of rows. Test approve, reject and timeout.

5

Add failure handling and the AI log (15 min)

Retry on the model node, error branch to the human queue, append-row to the AI log.

6

Evaluate (25 min)

Fifteen or more golden items with agreed labels. Score. Make one prompt change, re-score, record in the changelog.

7

Write the risk assessment and dashboard (15 min)

One page and one tab, using the Module 11 and 12 templates. Include the kill switch.

8

Prepare the presentation (10 min)

Five slides: problem, map, live run, controls, numbers and next steps. Rehearse once against the clock.

Deliverable

The capstone-<your name> folder with all seven deliverables, submitted before Session 14, plus your completed peer-review checklist for the capstone you were assigned.

Knowledge check

Pick one answer per question, then check your score. These mirror the style of the final exam.

1. A capstone has an excellent map and build but no approval gate on its outgoing customer email. What is the maximum mark?

100% 75% 50% It is not marked

Why: An ungated external action would not be safe to deploy, so the capstone is capped at 50% regardless of other components.

2. Which component carries the most weight in the rubric?

Workflow map Build Controls Presentation

Why: Build is 30%, controls 25%, map 20%, monitoring 15%, presentation 10%.

3. What is the minimum number of golden-set items for the capstone?

5 15 50 200

Why: The rubric asks for at least 15 agreed-label items, enough to give a meaningful accuracy figure within the time budget.

4. Your future-state map takes 45 minutes to draw. What does that tell you?

You are being thorough The process is too big for the capstone time budget You need more tools Nothing

Why: The guidance is that a map taking over 30 minutes signals a process that will not be finished to the standard in the time available.

5. What should the five-minute presentation lead with?

The list of nodes used The problem, the numbers and the controls The vendor's pricing A history of AI

Why: Markers and colleagues remember the process and its safeguards, not the tool names.

Self-check

Answer in your own words first, then open the model answer.

1. Write your capstone's one-line goal, trigger, final output and three tools.

Any answer that names all four concretely, for example: goal = handle leave requests without manual copying; trigger = form submission; output = calendar entry and confirmation email; tools = Google Forms, n8n, Google Calendar, Gmail.

2. Which of the seven deliverables would you produce first, and why?

The maps, because scoping and tiering decide everything that follows, and because a map that takes too long reveals a process that is too big before any build time is spent.

3. Name three items on the peer-review checklist that relate to Module 10 and 11 controls.

Every Tier 3+ action gated; approval message shows the full action; safe timeout; dedicated limited credential; audit log with approve, reject and timeout rows.

Summary

Key takeaways

Small and complete beats large and unfinished: three tools, one AI step, every gate, real numbers.
The rubric weights build 30, controls 25, map 20, monitoring 15, presentation 10. No gate on an external action caps the mark at 50%.
Seven deliverables: maps, export, approval evidence, golden set and score, dashboard, risk assessment, slides.
Review a peer with the checklist; the most common failures are oversized scope, empty approval messages, personal credentials and missing timeouts.

Further reading: A larger multi-agent capstone from the Computing Systems course · DevOps Lab capstone