One document for the whole system, every decision defended, plus the review you present and defend live. The artifact you take to work on Monday.
What this is
Six weeks converge into one artifact.
Nothing new to invent, your four projects are the sections. The capstone is the connective tissue: showing every decision fits with the others, then defending it under real questions. Pairs with Lab 6.
💡 In plain English
Assemble the car and drive it in front of a panel. You built the engine, the brakes, and the crash test in Projects 1 to 4. Now you show they are one machine, and answer "why is this safe on the road?"
Do this, in order
Assemble, then defend.
1
Stitch your four projects into the seven sectionsProblem, boundary, context, cost/reliability, architecture/security, evals/observability, trade-offs. Write the connective sentences.
2
Draw the system on one pageBoundary, retrieval, model core, guardrails, and observability around it. If it doesn't fit on a page, simplify.
3
Write the trade-offs sectionThree big decisions: what you chose against, and what risk remains.
4
Pre-answer the hard questionsWhy not just an agent, what's your worst failure, what under injection, how do you know it works.
5
Prepare the five-minute walkthroughProblem, the bar with its why, the shape, your two hardest calls (trade-off + trigger), the risk you kept.
Fill this in
The document skeleton.
Problem & scope
what it does, for whom, at what bar: …
Boundary
code vs model: … (P-boundary / S2)
Context & retrieval
what the model sees: … (Project 1)
Cost, latency, reliability
ceilings + how it fails safely: … (Project 2)
Architecture & security
pattern + defenses: … (Project 3)
Evals & observability
how you know it works: … (Project 4)
Trade-offs & risks
chosen against + residual: …
How it's marked
Coherent & defended vs a pile of docs.
✓ Strong
Seven sections that connect, plus a system that fits on one page.
Hard questions pre-answered; trade-offs and residual risk named.
A crisp five-minute walkthrough that goes straight to the two hard calls.
✕ Weak
Four projects stapled together with no connective tissue.
No number, or security bolted on after the fact.
A beautiful design with no way to know it works.
Wrapping up
Present it. Defend it. Ship it.
The finished document, the one-page diagram, and the presentation you give (plus the notes you take from the room). Then apply one decision from it to a real system at work this week.
Deliverable
The seven-section doc, the one-page diagram, and your review takeaways. Due Thu Aug 20.
submit on Maven
Stuck?
Use the S11 / Week 6 checklist as the outline. Most of the content is already in Projects 1 to 4.
rewatch S11 + S12
For the instructorHow to teach thisclick to open ▾
Reassure them there is almost nothing new to write, it is assembly and coherence. The clinic's whole purpose is to surface the gaps now, with you, not in front of the group or in production. Run the walkthrough as a real design review and ask the four hard questions.
End by coming full circle to their Week-1 goal, the moment their LLM feature broke, and show that this design has the answer in it. That is the emotional close of the entire course: not another demo, a system that holds.