Week 13 — Buffer

Why this week: twelve weeks rarely go to plan. This week exists so that one week that didn't click gets a proper second go, rather than being skipped. If everything went well, use it to go deeper on whatever interviewers keep asking about, or take the week off before the next project.

By Friday you have: one of three things, chosen on Monday morning: a repeated week, finished properly this time; a deeper piece of work tied to the questions you're being asked; or a rested head and a short plan for what comes next.

Main online course: whichever course belongs to the week you repeat. Nothing new.

A public holiday this week

The Ministry of Manpower lists Hari Raya Puasa 2027 on Wednesday 10 March. Plan four days, not five.

Concepts

Repeating is not failing. A week "didn't click" when you could follow the steps but couldn't explain the decisions, or when the deliverable works but you'd struggle to rebuild it from a blank file. Interviewers find that gap quickly. A second pass, with the first attempt's mistakes in front of you, usually takes half the time and sticks.

Depth that matches demand. By now you've had interviews and screening calls. The questions they ask are better evidence about what to learn next than any syllabus.

Monday — Choose

Build:

1. Read your notes: notes/progress.md (one line per week), every Friday check-in reply from Claude, and notes/interview-answers.md. Mark each week as solid, shaky or missing.

2. Pick one option, using this test:

   • Repeat a week if any week is marked missing, or shaky in a skill your target roles name in their job ads. If two qualify, pick the one closest to a role you're interviewing for now.
   • Go deeper if nothing is missing, and the same topic has come up in two or more interviews or screening calls.
   • Take the week off if every week is solid, and nothing keeps coming up in interviews.

3. Write the choice, and the reason, as one line in notes/progress.md. Then write this week's checklist into your README, as on every Monday.

Stuck choosing?

Paste your notes/progress.md into the Friday check-in prompt with this extra line: "Which single week should I repeat, and why? Name the one deliverable that would prove it clicked this time." Then decide for yourself; it's your evidence, not Claude's.

Option A — Repeat a week

Redo the week's page from its Monday, in a new branch of your repo, without looking at your first attempt until you're stuck. Then compare. Pages: week 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11.

The weeks people most often repeat, and what "clicked" looks like the second time:

Week	It clicked when you can...
2–3, agents	write the tool loop from a blank file, and explain each stopping rule
4, RAG	explain why hybrid search beat vector search alone on your data, with the recall numbers
5, error analysis	label 20 new messages and place each error in your categories without hesitating
6, PySpark and graphs	build one per-account window feature in Spark without copying, and explain what a connected component means for scam syndicates
7, judges and CI	say how well your judge agrees with your labels, and what you'd do if it dropped
9, backend	draw the request path from the browser to the database, and say where each secret lives
11, production	explain where the time and money go in one trace, without looking

Done when: the week's deliverable checklist is ticked again, in your own words, and the "clicked" line above is true.

Option B — Go deeper, led by interviews

Pick the topic that came up most, and spend the week on one piece of work you can show. Some starting points, matched to common interview questions:

They asked...	A week's deeper dive	Read first
"How do you know your agent works?"	Grade the agent's tool choices on 50 traces, and add the grader to CI	Demystifying evals for AI agents
"How would you trust an LLM judge?"	Measure judge agreement on 100 fresh labels, and try two rubric changes	Using LLM-as-a-Judge
"What stops prompt injection?"	Red-team the tools that read web pages, and reduce what each can reach	The lethal trifecta
"Why not just use a framework?"	Rebuild one feature on the framework from week 3, and compare code, cost and speed	Building effective agents
"How do you manage context?"	Measure how answer quality changes as retrieved context grows, and pick a limit	Effective context engineering
"How does this scale?"	Load-test the API with 50 parallel checks, and read the p95 and the Cloud Run instance count	Reducing latency
"Tell me about fraud data at scale"	Extend the week 6 graph features, and measure how much they lift the fraud model	PySpark: Getting started

Whatever you pick: one short write-up at the end, with a number in the first paragraph, linked from your README.

Done when: the work is in the repo, the write-up is linked, and you've added a better answer to notes/interview-answers.md for the question that started it.

Option C — Take the week off

If every week is solid, rest. Your applications are out, the product is live, and the next project will go better with a clear head. Before you stop:

1. Check the live service works, and that the Console, Google Cloud and AWS budgets and spend limits are still in place. Delete anything on AWS you're not using.
2. Write five lines in notes/next.md: what you'd build next, and which role it's aimed at.
3. Then close the laptop.

Interview practice, for every option

Whatever you choose, spend about an hour a day on this. It's the most direct use of the course's output.

1. Open notes/interview-answers.md. Pick three questions at random each day.
2. Answer each one aloud, timed: two minutes, with one number and one trade-off in each answer. Record yourself if you can bear it.
3. Rewrite the written answer where the spoken one was better, or where you got stuck.
4. On Thursday, ask Claude to play the interviewer for your target role: paste a real job posting and your case study, and ask for six questions, one at a time, with follow-ups on your weakest answers.

Friday — Close the course

Demo day (9:00–12:00):

1. Record a 3-minute video of whatever this week produced: the repeated deliverable, the deeper dive, or (option C) a calm tour of the live product.
2. Write five lines on what you learned this week.
3. Send both to Claude with the Friday check-in.

Afterwards: if you'd like the badge, the Cloud Practitioner exam is described on the cloud side track.

Deliverable checklist

• A one-line choice with its reason in notes/progress.md
• Option A: the repeated week's checklist ticked again; or option B: the deeper dive and its write-up; or option C: budgets checked and notes/next.md written
• Interview practice on at least four days, with notes/interview-answers.md improved
• Demo video, five lines, and the Friday check-in sent

If you have time left: turn off anything you don't need running. A product with a live link is worth keeping up; a forgotten AWS load balancer isn't.