Everything converges. Combine your five labs into one system behind a single entry point, write the seven-section design document, run your eval harness against the whole thing, and defend it in five minutes.
~4–6 hoursyour Week 1-5 builds
What you'll build
Not five labs. One system that holds.
A single serve(request) that routes, retrieves grounded context, calls the model inside a budget with a cache and a fallback ladder, validates and guards the output, and is fully traced and eval-gated. Then the document that defends every decision.
💡 In plain English
You have built the engine, the brakes, the fuel gauge, the seatbelts, and the crash test. This week you assemble the car, drive it, and explain to a panel why it is safe to put on the road.
The shape
Five labs, one entry point.
flowchart LR
IN["request"] --> R["router Lab 1"]
R --> CTX["hybrid retrieval Lab 2"]
CTX --> CALL["budgeted call + cache + fallback Lab 3"]
CALL --> AG["guardrails + approval Lab 4"]
AG --> OUT["validated answer"]
EVAL[("eval harness + traces Lab 5")] -.-> CALL
EVAL -.-> AG
classDef ok fill:#D6F5E3,stroke:#1F2937,color:#0E1726;
classDef ops fill:#DCEBFE,stroke:#1F2937,color:#0E1726;
class R,CTX,CALL,AG,OUT ok; class EVAL ops;
Do this, in order
The steps.
1
Wire the five labs into one entry pointRouter, then retrieval, then the budgeted/cached/fallback call, then guardrails and validation. One serve(request).
2
Run it end to end on a real requestA genuine input from your use case, all the way to a validated answer, within your budget.
3
Point your eval harness at the whole systemRun the golden set (including the safety case) against the integrated system. Green means shippable.
4
Write the seven-section design documentProblem, boundary, context, cost/reliability, architecture/security, evals/observability, trade-offs. Mostly assembled from Projects 1-4.
5
Draw the system on one pageIf it does not fit on a page, simplify until it does.
6
Prepare the five-minute walkthroughProblem, the bar with its why, the shape, your two hardest calls (trade-off + trigger), and the risk you kept.
You're done when
Acceptance criteria.
The system runs end to end on a real request, within budget.
The design document has all seven sections, and they connect.
The eval harness runs green on the integrated system, safety case included.
The whole system fits on one page.
You can give the five-minute walkthrough: problem, bar, shape, two hard calls, the risk you kept.
Go further
Stretch goals.
Add the three dashboards (cost, latency, quality) with real alert lines.
Turn one real failure into a new eval case live during the demo, the flywheel in action.
Load-test the fallback: kill the primary model mid-demo and show it degrade, not break.
Write the one-paragraph "what I would do with two more weeks", the honest residual-risk note.
Wrapping up
Present it. Defend it. Ship it.
Hand in the integrated system, the seven-section document, the green eval run, and your walkthrough. This is your Project 5 (System Design Document + Review), the artifact you take to work on Monday.
Deliverable
The system, the doc, the one-page diagram, and the green eval run.
this is Project 5
Stuck?
Integrate two labs at a time. A system with the router and the budgeted call wired is already defensible; add the rest.
rewatch S11
For the instructorHow to run this labclick to open ▾
The magic is that there is almost nothing new to build, it is assembly and defence. Reassure the ones who feel behind: a system with three of the five labs wired, honestly documented, beats a "complete" system nobody can explain. The document and the walkthrough are graded, not line count.
Run the five-minute walkthrough as a real design review. Ask the four hard questions from the checklist, let the group find the gaps, and end by pointing back to their Week-1 goal, the moment their LLM feature broke in production, and show that this system has the answer in it. That full-circle is the close of the whole course.