Start here — Session 0
This page takes you from a fresh laptop to a running course notebook. It covers the ten-session arc, the tools, the student kit, the five accounts, the domain the CDN session needs, your SSH key, the safe-data policy, the assessment and a readiness checklist. Work through it in order; each section ends with the result that proves it worked.
Three anchors hold everything else up:
- Accounts — one provider account per lab, each with two-factor authentication and a spend guardrail.
- The kit —
cloud-tech-kit.zip, which installs a pinned Python 3.14 environment and fetches the course data. - Safe data — only the course's synthetic media goes to any cloud service.
How to read this page. A command in a bash or PowerShell block is followed
immediately by two things: a sentence or two saying what the command does and why
it is that command, and the output that proves it worked. Commands marked Live
account need a real provider account. Their expected result is stated exactly;
the instructor's sanitized run is recorded in the course runbook (issue #112).
Enter every placeholder written like <this> with your own value. Your instructor
fills the three owner-only placeholders before publication.
The course in one page
Ten four-hour sessions, each with a research answer and a lab, then a 30-point capstone. The campus media unit in Chiang Mai — 40 short videos a month and a 6,000-asset image and 3D catalogue — is the scenario that every session reuses. Each session is worth 7 weekly points (4 research + 3 lab), for 70; the capstone is the other 30.
| # | Session | What you must have ready before it | Points |
|---|---|---|---|
| 01 | Cloud foundations | DigitalOcean account with 2FA and a backup method, instructor as Resource Viewer, $10 spend alert, SSH key, doctl |
7 |
| 02 | Performance and latency | A root domain with an Active Cloudflare zone; a DigitalOcean token with droplet:update; dig |
7 |
| 03 | Storage, RAID and ZFS | The DigitalOcean account and token; a Block Storage volume | 7 |
| 04 | Networking and delivery | The Active Cloudflare zone, R2 enabled, Node ≥ 22.12, wrangler |
7 |
| 05 | Containers and the Nextcloud NAS | The S4 zone and DNS token; doctl, ssh, scp, dig, curl, jq, rclone |
7 |
| 06 | Economics and vendor lock | The Linode/Akamai account, payment method, 2FA and a per-service token; linode-cli, aws, jq; the OpenRouter account and key |
7 |
| 07 | Security and identity | The S1 SSH key; DigitalOcean and Cloudflare tokens; doctl, wrangler, aws, jq, openssl |
7 |
| 08 | Sensitive data and managed AI | The OpenRouter key from S6; the S7 risk-register rows; Pillow 11.3.0 from the kit | 7 |
| 09 | LLM inference and fine-tuning | A phone-verified Kaggle account, or Apple silicon with the kit's S9 group; internet on for Kaggle's install cell | 7 |
| 10 | Image ML on a GPU + capstone | A phone-verified Kaggle account; the kit's S10 group or the light install; the fetched capstone pack | 7 |
What the labs actually rent. A Droplet is a virtual machine with a disk and a public IPv4 address. A Worker is a managed runtime for code. An inference API is a managed model on someone else's hardware. The useful question in every session is not "which layer?" but who owns the operating system, the patching and the failure mode. Session 1 sets the account up; Sessions 2–8 spend from it; S9 and S10 move the heavy compute to Kaggle or the faculty GPU.
สรุปภาษาไทย
หน้านี้พาผู้เรียนจากเครื่องใหม่ไปจนถึงการรันโน้ตบุ๊กของรายวิชา ครอบคลุมลำดับสิบเซสชัน เครื่องมือ ชุดนักเรียน (คิต) ห้าบัญชี โดเมนสำหรับเซสชัน CDN กุญแจ SSH นโยบายข้อมูลปลอดภัย การประเมิน และรายการตรวจความพร้อม แต่ละเซสชันมีคะแนน 7 คะแนน (วิจัย 4 + แล็บ 3) รวม 70 คะแนน และแคปสโตน 30 คะแนน ต้องมีบัญชี DigitalOcean ที่เปิด 2FA และวิธีสำรอง ตั้งอาจารย์เป็น Resource Viewer ตั้งแจ้งเตือน 10 ดอลลาร์ และมีกุญแจ SSH ก่อนเซสชัน 1 ก่อนเซสชัน 2 ต้องมีโดเมนระดับรากและโซน Cloudflare ที่ Active ก่อนเซสชัน 6 ต้องมีบัญชี Linode/Akamai และก่อนเซสชัน 9–10 ต้องมีบัญชี Kaggle ที่ยืนยันเบอร์โทรแล้ว คำสั่งที่ทำเครื่องหมาย Live account ต้องใช้บัญชีจริง ผลลัพธ์ที่คาดหวังระบุไว้ชัดเจน และผู้สอนบันทึกการรันที่ปกปิดข้อมูลแล้วในรันบุ๊ก中文总结
本页带学习者从一台新电脑走到能运行课程 notebook,涵盖十节课的顺序、工具、学生工具包(kit)、五个账号、用于 CDN 一课的域名、SSH 密钥、安全数据政策、评估以及准备清单。每节课 7 分(研究 4 + 实验 3),合计 70 分,另有 capstone 30 分。在第 1 课之前,必须有一个启用 2FA 并设置好备份方式的 DigitalOcean 账号、把教师设为 Resource Viewer、设置 10 美元提醒,并准备好 SSH 密钥。第 2 课之前,必须有一个根域名和一个处于 Active 状态的 Cloudflare zone。第 6 课之前,必须有 Linode/Akamai 账号;第 9–10 课之前,必须有一个已完成手机验证的 Kaggle 账号。标记为 Live account 的命令必须使用真实账号,预期结果已明确写出,教师已在 runbook 中记录脱敏后的运行过程。Tools, per operating system
The course runs everything inside a virtual environment: a private folder of
Python packages created from the kit's lockfile. uv builds that environment and
downloads Python 3.14; git is how you get and version files; the rest are the
provider command-line interfaces the labs use. You do not need all eight tools on
day one, but installing them now means no lab starts with a missing command.
| Tool | Needed for | What it is |
|---|---|---|
uv |
every session | installs Python 3.14 and the pinned packages into .venv |
git |
every session | version control; also how you copy sample projects |
| VS Code + Jupyter extension | running notebooks | the editor and the notebook runtime |
doctl |
S1–S3, S5–S7 | DigitalOcean's CLI |
wrangler |
S4, S6, S7 | Cloudflare Workers and R2 CLI, run from Node |
linode-cli |
S6 | Linode/Akamai's CLI |
aws |
S6, S7 | the S3-compatible client for Linode Object Storage |
jq |
S1, S5–S8 | reads and filters JSON output |
Your machine baseline. macOS 13 or newer on Apple silicon, Windows 11, or a current Linux distribution. You need about 3 GB of free disk for the full kit (under 1 GB for the light install), 4 GB of RAM, and a network connection for the first run. Intel Macs work for S1–S8; S9 and S10 use Kaggle or the faculty GPU instead of local MLX.
macOS. Every tool comes from Homebrew. Install Homebrew
first if brew --version fails; the installer asks for your password and explains
each step before it runs. brew --version prints a version line such as
Homebrew 4.x.
/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)"
This runs Homebrew's official installer. It is a network install, so read the
script before answering the prompt; it writes under /opt/homebrew on Apple
silicon. Proof: brew --version prints the Homebrew version.
| Tool | Command | What it does | Proof |
|---|---|---|---|
| uv | brew install uv |
installs the Python and package manager | uv --version prints uv 0.13.x or newer |
| git | brew install git |
installs version control | git --version prints git version 2.x |
| jq | brew install jq |
installs the JSON filter | jq --version prints jq-1.8.x |
| doctl | brew install doctl |
installs the DigitalOcean CLI | doctl version prints a version and the API endpoint |
| aws | brew install awscli |
installs the S3-compatible AWS CLI | aws --version prints aws-cli/2.x |
| linode-cli | brew install linode-cli |
installs the Linode/Akamai CLI | linode-cli --version prints linode-cli v5.x |
| Node | brew install node |
installs Node, which runs Wrangler | node --version prints v22.x or newer |
| Wrangler | npm install -g wrangler |
installs the Cloudflare CLI globally | wrangler --version prints 4.x |
| VS Code | brew install --cask visual-studio-code |
installs the editor | code --version prints 1.1xx.x |
| Jupyter extension | code --install-extension ms-toolsai.jupyter |
adds notebook editing and kernel support to VS Code | the command prints Extension 'ms-toolsai.jupyter' was successfully installed. |
| Python extension | code --install-extension ms-python.python |
adds Python language and interpreter selection | the command prints Extension 'ms-python.python' was successfully installed. |
Windows 11. Use winget for the packaged tools and the built-in Windows
OpenSSH client for ssh. winget --version prints the App Installer version.
If a package id has changed, winget search <name> shows the current id.
winget --version
This proves the Windows Package Manager is present. Expected output is a version
line such as v1.x; if the command is not found, update App Installer from
the Microsoft Store.
| Tool | Command | What it does | Proof |
|---|---|---|---|
| OpenSSH | Add-WindowsCapability -Online -Name OpenSSH.Client~~~~0.0.1.0 |
installs the built-in OpenSSH client when it is absent; run it in an Administrator PowerShell | ssh -V prints OpenSSH_for_Windows_9.x or newer |
| uv | winget install --id astral-sh.uv -e |
installs uv | uv --version prints uv 0.13.x or newer |
| git | winget install --id Git.Git -e |
installs Git for Windows, including ssh-keygen |
git --version prints git version 2.x |
| jq | winget install --id jqlang.jq -e |
installs the JSON filter | jq --version prints jq-1.8.x |
| Node | winget install --id OpenJS.NodeJS.LTS -e |
installs Node LTS, which runs Wrangler | node --version prints v22.x or newer |
| Wrangler | npm install -g wrangler |
installs the Cloudflare CLI globally | wrangler --version prints 4.x |
| linode-cli | uv tool install linode-cli |
installs the Linode CLI in an isolated uv-managed environment | linode-cli --version prints linode-cli v5.x |
| aws | winget install --id Amazon.AWSCLI -e |
installs the AWS CLI | aws --version prints aws-cli/2.x |
| VS Code | winget install --id Microsoft.VisualStudioCode -e |
installs the editor | code --version prints 1.1xx.x |
| Jupyter extension | code --install-extension ms-toolsai.jupyter |
adds notebook support | the command prints Extension 'ms-toolsai.jupyter' was successfully installed. |
| Python extension | code --install-extension ms-python.python |
adds Python support | the command prints Extension 'ms-python.python' was successfully installed. |
doctl has no winget package. Download the doctl-<version>-windows-amd64.zip
file from the doctl releases page,
extract it, move doctl.exe to a folder on your PATH (for example
%USERPROFILE%\bin), and open a new PowerShell window.
doctl version then prints doctl version 2.x and
API endpoint: https://api.digitalocean.com/, which proves the binary is on
PATH.
Linux. Use your distribution's package manager. The example is Debian/Ubuntu
(apt); Fedora uses dnf, Arch uses pacman. sudo asks for your password and
installs system-wide. Proof for each row is the same version command as above.
| Tool | Command | What it does | Proof |
|---|---|---|---|
| git, jq, aws | sudo apt update && sudo apt install -y git jq awscli |
installs three packaged tools | git --version, jq --version, aws --version |
| uv | curl -LsSf https://astral.sh/uv/install.sh | sh |
runs uv's official installer into ~/.local/bin |
uv --version prints uv 0.13.x or newer |
| doctl | sudo snap install doctl |
installs the DigitalOcean CLI from Snap | doctl version |
| Node | sudo snap install node --classic --channel=22 |
installs Node 22 from Snap, which Wrangler needs | node --version prints v22.x or newer |
| Wrangler | npm install -g wrangler |
installs the Cloudflare CLI globally | wrangler --version prints 4.x |
| linode-cli | uv tool install linode-cli |
installs the Linode CLI in an isolated uv environment | linode-cli --version prints linode-cli v5.x |
| VS Code | sudo snap install code --classic |
installs the editor | code --version prints 1.1xx.x |
| Jupyter extension | code --install-extension ms-toolsai.jupyter |
adds notebook support | the command prints Extension 'ms-toolsai.jupyter' was successfully installed. |
| Python extension | code --install-extension ms-python.python |
adds Python support | the command prints Extension 'ms-python.python' was successfully installed. |
uv tool install writes the executable to ~/.local/bin (%USERPROFILE%\.local\bin
on Windows). Open a new terminal so the shell sees it; if linode-cli --version
still reports "command not found", add that folder to PATH.
Prove every install at once. Run these commands in the order shown.
Each prints a version line; a non-zero exit or "command not found" means that
tool is not on PATH yet. Versions move, so the proof is the line's shape, not
the exact number. These were verified on macOS on 2026-11-15:
| Command | Expected output (shape) |
|---|---|
uv --version |
uv 0.13.8 (3d97a6aad 2026-11-12 aarch64-apple-darwin) |
git --version |
git version 2.56.0 |
code --version |
1.143.1, then a commit hash, then arm64 (or x64) |
doctl version |
doctl version 2.4.0-release, then API endpoint: https://api.digitalocean.com/ |
wrangler --version |
4.90.0 |
linode-cli --version |
linode-cli v5.70.0, then Built from spec version 4.229.1 |
aws --version |
aws-cli/1.44.73 Python/3.11.9 Darwin/27.0.0 botocore/1.42.83; the Homebrew 2.x line reads aws-cli/2.x Python/3.x <platform> |
jq --version |
jq-1.8.2 |
ssh -V |
OpenSSH_10.6p1, OpenSSL 4.1.0 23 Oct 2026 |
node --version |
v22.19.0 |
The course was verified with doctl 2.4.0 and Wrangler 4.90.0. An older doctl
may not know doctl projects or doctl balance get; an older Node makes
Wrangler refuse to start. Upgrade before the lab rather than during it.
สรุปภาษาไทย
เครื่องมือทั้งหมดติดตั้งผ่านตัวจัดการแพ็กเกจของระบบปฏิบัติการ macOS ใช้ Homebrew, Windows 11 ใช้ `winget` ร่วมกับ Windows OpenSSH ที่มีมาในระบบ และ Linux ใช้ตัวจัดการแพ็กเกจของดิสโทร (ตัวอย่างเป็น apt) เครื่องมือที่ต้องมีคือ `uv` สำหรับสร้างสภาพแวดล้อม Python 3.14, `git`, VS Code พร้อมส่วนขยาย Jupyter และ Python, `doctl`, `wrangler` (ต้องมี Node 22 ขึ้นไป), `linode-cli`, `aws` และ `jq` หลังติดตั้งให้รันคำสั่งตรวจเวอร์ชันทีละตัว เช่น `uv --version`, `git --version`, `doctl version`, `wrangler --version` ผลลัพธ์ที่ถูกต้องคือบรรทัดเวอร์ชันและรหัสออก 0 ถ้าเจอ command not found แปลว่าเครื่องมือยังไม่อยู่ใน PATH ต้องเปิดเทอร์มินัลใหม่หรือเพิ่มPATH ก่อนเริ่มแล็บ ต้องมีดิสก์ว่างประมาณ 3 GB สำหรับคิตเต็ม หรือต่ำกว่า 1 GB สำหรับการติดตั้งแบบเบา และมีแรม 4 GB中文总结
所有工具都通过操作系统的包管理器安装:macOS 使用 Homebrew,Windows 11 使用 `winget` 配合系统自带的 Windows OpenSSH,Linux 使用发行版的包管理器(示例为 apt)。必需的工具包括:用于创建 Python 3.14 环境的 `uv`、`git`、带 Jupyter 和 Python 扩展的 VS Code、`doctl`、`wrangler`(需要 Node 22 及以上)、`linode-cli`、`aws` 和 `jq`。安装后逐个运行版本检查命令,例如 `uv --version`、`git --version`、`doctl version`、`wrangler --version`;正确结果是版本行和退出码 0。如果看到 command not found,说明工具还不在 PATH 中,需要打开新的终端或把路径加入 PATH。开始实验前,完整 kit 需要约 3 GB 可用磁盘空间,轻量安装则需要低于 1 GB,并需要 4 GB 内存。The student kit
The kit is a single archive, cloud-tech-kit.zip, that contains the course's
pinned environment (pyproject.toml and uv.lock), the setup and start scripts,
VS Code settings, and every file the site publishes for the ten sessions. It
mirrors the site's folder shape, so a notebook that reads data/cache_sample.csv
or ../../assets/synthetic-media/photo-01.jpg resolves the same way inside the
kit. It does not contain the data bytes: the first run downloads them and
checks each file's SHA-256 against data-manifest.json. That keeps the archive
small and makes a corrupt download fail loudly instead of silently.
The kit installs into a .venv folder inside itself. Nothing is installed
system-wide, and deleting the kit folder removes the environment. The kit pins
Python 3.14 and every package; do not edit uv.lock, and do not run a bare
uv run after setup, because that re-resolves the default groups and would
reinstall the S9/S10 packages a light install deliberately skipped.
Download and verify
Download two files from the course site into the same folder, for example
~/Downloads:
SHA256SUMS holds the archive's digest. Recomputing the digest and comparing it
proves the download arrived complete and unmodified.
macOS
cd ~/Downloads
shasum -a 256 -c SHA256SUMS
shasum -c recomputes the archive's SHA-256 and compares it with the line in
SHA256SUMS. Proof: it prints cloud-tech-kit.zip: OK. Any other line means the
download is incomplete; delete both files and download again.
Linux
cd ~/Downloads
sha256sum -c SHA256SUMS
The GNU tool performs the same comparison. Proof: cloud-tech-kit.zip: OK.
Windows (PowerShell)
cd $HOME\Downloads
Get-FileHash .\cloud-tech-kit.zip -Algorithm SHA256 | Format-List
Get-Content .\SHA256SUMS
The first command prints the archive's hash under Hash; the second prints the
expected line. Proof: the two 64-character values match. If they do not, delete
and download again.
Unpack and install
The setup script performs four steps and prints each one: it checks uv, installs
Python 3.14, creates .venv from uv.lock, and downloads the data. It stops with
the per-OS install command if uv is missing; it never installs uv itself, so
you can see and approve the installer you run.
macOS / Linux
unzip -q cloud-tech-kit.zip
cd cloud-tech-kit
ls
unzip creates the cloud-tech-kit/ folder and cd moves into it. Proof: ls
lists README.md, pyproject.toml, uv.lock, setup.sh, start-jupyter.sh and
weeks/.
bash setup.sh
bash runs the setup script without needing the executable bit. The long step is
the package download: about 120 MB for the notebooks group, plus the S9 and S10
groups that apply to your platform. Proof: the script prints == Step 1 of 4,
== Step 2 of 4, == Step 3 of 4, == Step 4 of 4 and finally == Done; the
fetch ends with a summary: line such as
summary: 197 downloaded, 0 skipped, 197 files, 3,851,590 bytes verified.
uv run --no-sync python -V
--no-sync means "use the .venv exactly as setup built it"; without it, uv
would re-sync the default groups first. Proof: it prints Python 3.14.x.
Windows (PowerShell)
Expand-Archive -Path .\cloud-tech-kit.zip -DestinationPath .
cd .\cloud-tech-kit
dir
Expand-Archive unpacks the archive into the current folder. Proof: dir shows
the same files as the macOS listing above.
.\setup.ps1
If PowerShell blocks local scripts, run it for this one invocation with the policy
relaxed: powershell -ExecutionPolicy Bypass -File .\setup.ps1. Proof: the same
four step headings and == Done, and the fetch's summary: line.
uv run --no-sync python -V
Proof: Python 3.14.x.
The light install. The S9 group is Apple-silicon MLX (several GB); the S10 group is PyTorch. If you will run S9 on Kaggle and S10 on Kaggle or the faculty GPU, skip both. Everything else — JupyterLab, S1–S8 — still installs.
bash setup.sh --no-group s09 --no-group s10
--no-group leaves the named dependency group out of the resolution. On Windows
the same flag works with .\setup.ps1 --no-group s09 --no-group s10. Proof: setup
still ends with == Done, JupyterLab and pip are present, and
uv run --no-sync python -c "import torch" fails with ModuleNotFoundError
because PyTorch was not installed. The choice sticks: setup and the start script
both use --no-sync, so they do not pull the groups back in.
The script is a wrapper around four commands. If you need to rebuild the
environment by hand, the light install itself is
uv sync --no-group s09 --no-group s10 --python 3.14: uv sync reads the lock
and creates .venv, --python 3.14 fixes the interpreter, and the two
--no-group flags skip MLX and PyTorch. Proof is the same — .venv exists and
uv run --no-sync python -V prints Python 3.14.x.
The first-run fetch
Step 4 downloads every file listed in data-manifest.json into the same relative
path. Each file is hashed before it is accepted, so a captive-portal page or a
cut-off transfer stops the run with exit code 2 and leaves the previous file
untouched. The fetch is safe to repeat: a file whose hash already matches is
skipped.
uv run --no-sync python fetch_data.py --all
This downloads the week data/ folders, the synthetic-media pack and the capstone
samples. Proof on a first run: one downloaded line per file, ending in
summary: 197 downloaded, 0 skipped, 197 files, 3,851,590 bytes verified. Run it
again and the proof becomes summary: 0 downloaded, 197 skipped, ….
uv run --no-sync python fetch_data.py --session s02
Use --session to fetch one week's data only; it accepts s02, 2 or the folder
slug 02-performance-latency. It never fetches the media pack or capstone
samples. Proof: the summary: line lists only files under
weeks/02-performance-latency/data/.
Open a notebook — VS Code
- Open the kit folder: File → Open Folder…, choose
cloud-tech-kit. From a terminal wherecodeis installed,code .does the same. - Accept the recommended extensions when VS Code offers them. The kit recommends
ms-python.pythonandms-toolsai.jupyter; if you declined, the twocode --install-extensioncommands in the Tools section install them. - Select the interpreter: View → Command Palette → Python: Select
Interpreter → Enter interpreter path →
.venv/bin/python(Windows:.venv\Scripts\python.exe). Proof: the status bar showsPython 3.14.x ('.venv': venv). - Open
weeks/02-performance-latency/notebook.ipynbin the Explorer and press Run All. The kit pinsjupyter.notebookFileRootto${fileDirName}, so the notebook'sdata/resolves to the week folder, not the kit root. Proof: every cell runs and the outputs mention the CSV rows the notebook read.
Open a notebook — JupyterLab
bash start-jupyter.sh
The script changes to the kit root and runs uv run --no-sync jupyter lab, so
every relative data/ path resolves exactly as it does in VS Code. Proof: the
terminal prints a http://localhost:8888/lab?token=… URL and your browser opens
it. In the file browser open weeks/02-performance-latency/, then
notebook.ipynb, and use Run → Run All Cells. Stop the server with Ctrl-C,
then confirm with y.
On Windows the equivalent is .\start-jupyter.ps1. Both paths run the same
notebook against the same .venv, so the proof is identical: Run All reaches the
last cell without an error.
สรุปภาษาไทย
คิตคือไฟล์ `cloud-tech-kit.zip` ที่บรรจุสภาพแวดล้อม Python 3.14 แบบล็อกเวอร์ชัน ไฟล์ตั้งค่า สคริปต์ติดตั้ง และไฟล์ที่เว็บไซต์เผยแพร่สำหรับทั้งสิบเซสชัน โดยคงโครงสร้างโฟลเดอร์เหมือนเว็บไซต์ ข้อมูลจริงไม่ได้อยู่ในซิป แต่จะถูกดาวน์โหลดในรอบแรกและตรวจ SHA-256 กับ `data-manifest.json` ให้ดาวน์โหลดซิปและ `SHA256SUMS` ไว้โฟลเดอร์เดียวกัน ตรวจด้วย `shasum -a 256 -c SHA256SUMS` บน macOS หรือ `sha256sum -c` บน Linux หรือ `Get-FileHash` บน Windows ผลที่ถูกต้องคือ `cloud-tech-kit.zip: OK` จากนั้นแตกไฟล์ เข้าโฟลเดอร์ และรัน `bash setup.sh` หรือ `.\setup.ps1` สคริปต์ทำสี่ขั้นและจบด้วย `== Done` หลักฐานคือ `uv run --no-sync python -V` พิมพ์ `Python 3.14.x` และ fetch จบด้วยบรรทัด `summary:` การติดตั้งแบบเบาใช้ `--no-group s09 --no-group s10` เพื่อข้าม MLX และ PyTorch เปิดโน้ตบุ๊กได้สองทาง คือ VS Code (เลือก interpreter `.venv` แล้ว Run All) หรือ JupyterLab (`bash start-jupyter.sh` แล้วเปิดโน้ตบุ๊กจากโฟลเดอร์สัปดาห์) ทั้งสองทางใช้ `.venv` เดียวกันและต้องรันถึงเซลล์สุดท้ายโดยไม่มี error中文总结
kit 是文件 `cloud-tech-kit.zip`,内含锁定版本的 Python 3.14 环境、配置文件、安装脚本,以及网站为十节课发布的全部文件,并保持与网站相同的文件夹结构。真实数据不在压缩包内,而是在首次运行时下载,并用 `data-manifest.json` 校验 SHA-256。请把压缩包和 `SHA256SUMS` 下载到同一个文件夹,在 macOS 上用 `shasum -a 256 -c SHA256SUMS`、在 Linux 上用 `sha256sum -c`、在 Windows 上用 `Get-FileHash` 进行校验;正确结果是 `cloud-tech-kit.zip: OK`。然后解压、进入文件夹,并运行 `bash setup.sh` 或 `.\setup.ps1`;脚本执行四个步骤,以 `== Done` 结束。证据是 `uv run --no-sync python -V` 输出 `Python 3.14.x`,并且 fetch 以 `summary:` 行结束。轻量安装使用 `--no-group s09 --no-group s10` 跳过 MLX 和 PyTorch。打开 notebook 有两种方式:VS Code(选择 `.venv` 解释器后 Run All)或 JupyterLab(`bash start-jupyter.sh`,再从对应周文件夹打开 notebook)。两种方式使用同一个 `.venv`,并且必须无错误地运行到最后一个单元格。Accounts, in the order the labs need them
Every lab rents something from a provider, so every lab needs an account that is safe to leave running. The ladder below is ordered by the session that first needs each account. Two rules apply to all of them:
- Two-factor authentication plus a backup method before you create anything billable. The account is the root of every other guardrail.
- Least privilege: create a scoped token for the labs, give it an expiry, and never reuse an account password. A token you cannot edit is safer than one that can do everything.
Budget. The course envelope is $25 per student, with $10 reimbursed
against a receipt through <reimbursement method>. DigitalOcean spend alerts are
email thresholds, not caps: the account keeps spending after the email arrives.
That is why every lab ends with a teardown and a bill read.
DigitalOcean — before Session 1
Why. The course's compute and block storage. A Droplet is billed for the hours
it exists, including while powered off; only destroying it stops the meter. The
lab size s-1vcpu-1gb is $6.00/month, $0.00893/hour, with 1,000 GiB of
transfer included. The course region is SGP1 (Singapore) — there is no
DigitalOcean region in Thailand, which is what makes the latency lesson real.
Signup path. Create the account at https://cloud.digitalocean.com/registrations/new. A payment card is required at signup; there is no student programme. Then:
- Account → Settings: turn on two-factor authentication and add a backup method (a second factor or recovery codes). Proof: the page shows Enabled and lists the backup method.
- Teams → Members: invite
<instructor email>with the Resource Viewer role. A Resource Viewer can see the account's resources but cannot change or delete them; that is the role the course needs for support. - Billing → Spend alerts → Add alert: set a $10 monthly budget and the percentage thresholds. Proof: the alert appears in the list. The alert emails you; it does not stop spending.
- API → Tokens → Generate New Token is Session 1's step, not this one. It creates the scoped token the labs use.
Verification. After Session 1 creates your token, run this command. It is marked Live account because it needs the token.
doctl account get --format Status,DropletLimit
doctl sends the token to the account endpoint and prints only the status and the
Droplet limit, so the account email does not appear. Proof: two columns, the first
active, the second a number. A 401 means the token is wrong; a 403 means a scope
is missing and you must create a new token, because DigitalOcean tokens cannot be
edited.
Screenshot rule. Capture the 2FA page with the backup method visible, and the spend-alert page. Never capture the account email, account UUID, project UUID or a token value.
Cloudflare — before Session 2
Why. R2 object storage and the CDN edge that carries thumbnails and video starts. The free plan is enough: R2 Standard includes 10 GB-month of storage, egress is free, and Workers Free allows 100,000 requests/day. Cloudflare Free cannot host a subdomain as its own zone, so the course uses a root domain that you own.
Signup path. Create a free account at https://dash.cloudflare.com/sign-up. No card is required for the Free plan. Add your root domain after the domain step, and take the Free plan. For Session 2's measurements you also need a zone that is Active, which means the nameserver change in the domain section has propagated.
Verification. After Session 4 creates a token, wrangler whoami proves the
CLI is authenticated and lists the accounts it can see. It is marked Live
account.
npx wrangler whoami
Wrangler runs through Node, so the npx form works even when it is not installed
globally. Proof: a table of account names and IDs followed by the token's
permission rows. Before you have a token, the proof is the dashboard login and the
Free plan shown under Workers & Pages → Plans.
Screenshot rule. Capture the zone in the Active state with the account ID covered; never capture the API token.
OpenRouter — before Session 6, used again in Session 8
Why. A single OpenAI-compatible API over many model endpoints. Session 6 makes one priced call; Session 8 applies the governance rule to the endpoint choice. OpenRouter's free-model allowance is 20 requests/minute and 50/day; buying at least $10 of credit raises it to 1,000/day. Credit purchases carry a 5.5% fee, charged at purchase.
Signup path. Create an account at https://openrouter.ai, enable 2FA, and add
$1–2 of credit. Session 6 creates the key and labels it course-s6-<id>. The
account and remaining credit stay for Session 8; the key is revoked at teardown.
Verification. This reads the account's credit and usage. It is marked Live account.
curl -s https://openrouter.ai/api/v1/credits -H "Authorization: Bearer $OPENROUTER_API_KEY" | jq .
/credits returns a data object with total_credits and total_usage; the
difference is what you can spend. Proof: valid JSON with both fields, and
total_usage at 0 before your first call.
Screenshot rule. Capture the key page with the secret covered and the key label visible, plus the top-up receipt. Never capture the key.
Linode / Akamai — before Session 6
Why. Session 6 moves the catalogue service to a second provider to price and measure portability. The comparison node is a Linode 2GB at $12.00/month; the smallest, the Nanode 1GB, is $5.00/month. Hourly billing rounds up to the nearest hour. Object Storage is $5.00/month for 250 GB and 1 TB of transfer; extra storage is $0.02/GB-month. New customers may have a $100 credit that requires a payment method.
Signup path. Create the account at https://cloud.linode.com, add a payment method, enable 2FA, then API Tokens → Create Personal Access Token. Give it read/write on Linodes, Volumes, Object Storage and Account, and an expiry of about 90 days.
Verification. Export the token for this shell only, then read the account. It is marked Live account.
export LINODE_CLI_TOKEN="<linode-token>"
linode-cli account view --format company,balance,balance_uninvoiced --text --no-headers
account view is a read-only call that proves the token. Proof: three columns —
company, balance and uninvoiced balance. Invalid Token means the value was
copied incompletely or has expired; create a new token rather than guessing.
Screenshot rule. Capture the token list with the token value hidden and the expiry visible. Redact the account ID.
Kaggle — before Session 9
Why. A free NVIDIA T4 ×2 (32 GB total) with 30 GPU-hours/week, used for S9's CUDA fine-tune and S10's GPU comparison. There is no charge and no card required, but Kaggle requires phone verification before it grants GPU access. Internet is off by default inside a Kaggle notebook; S9's install cell needs it on.
Signup path. Create an account at https://www.kaggle.com, verify the phone number under Settings → Phone Verification, then in a notebook choose Settings → Accelerator → GPU T4 ×2.
Verification. There is no laptop CLI. Inside a Kaggle notebook, run this cell:
!nvidia-smi
nvidia-smi asks the runtime which GPU it received. Proof: a table that names a
Tesla T4 and shows its memory. The Settings panel must also show GPU T4 ×2
and Internet on for the install cell.
Screenshot rule. Capture the accelerator panel and the nvidia-smi line; the
notebook is private, so no personal data leaves it.
The domain — before Session 2
See A domain before Session 2. It is a separate step because it is the one thing you buy rather than sign up for, and because the Cloudflare zone waits on it.
สรุปภาษาไทย
บัญชีทั้งหมดเรียงตามเซสชันที่ต้องใช้ครั้งแรก และทุกบัญชีต้องเปิดสองปัจจัย (2FA) พร้อมวิธีสำรองก่อนสร้างทรัพยากรที่คิดเงิน DigitalOcean ใช้สำหรับเครื่องเสมือนและบล็อกสตอเรจ ต้องมีบัตรตอนสมัคร ตั้งอาจารย์เป็น Resource Viewer ตั้งแจ้งเตือน 10 ดอลลาร์ต่อเดือน และสร้างโทเคนแบบจำกัดสิทธิ์ในเซสชัน 1 งบประมาณรายวิชาคือ 25 ดอลลาร์และคืนเงิน 10 ดอลลาร์ผ่านวิธีการที่ผู้สอนกำหนด Cloudflare ใช้ฟรีสำหรับ R2 และ CDN โดยมีพื้นที่ฟรี 10 GB-month และอีเกรสฟรี ต้องมีโซนที่ Active ก่อนเซสชัน 2 OpenRouter ใช้สำหรับเรียกโมเดลแบบเสียเงิน ต้องเติมเครดิตอย่างน้อย 10 ดอลลาร์เพื่อเพิ่มโควตารุ่นฟรีเป็น 1,000 ครั้งต่อวัน Linode/Akamai ใช้ในเซสชัน 6 ต้องเพิ่มวิธีชำระเงิน เปิด 2FA และสร้างโทเคนอายุประมาณ 90 วัน Kaggle ใช้ฟรีสำหรับ GPU T4 สองตัว สัปดาห์ละ 30 ชั่วโมง ต้องยืนยันเบอร์โทร และต้องเปิดอินเทอร์เน็ตในโน้ตบุ๊กเมื่อติดตั้งแพ็กเกจ คำสั่งตรวจสอบที่ทำเครื่องหมาย Live account ต้องใช้บัญชีจริงและผลลัพธ์ที่คาดหวังระบุไว้แล้ว中文总结
所有账号按首次需要它们的课程顺序排列,并且每个账号在创建计费资源之前都必须启用双重验证(2FA)并设置好备份方式。DigitalOcean 用于虚拟机和块存储,注册时需要银行卡;把教师设为 Resource Viewer,设置每月 10 美元的提醒,并在第 1 课创建受限权限的 token。课程预算是 25 美元,并通过教师指定的方式报销 10 美元。Cloudflare 免费用于 R2 和 CDN,提供 10 GB-month 的免费存储和免费 egress;第 2 课之前必须有一个处于 Active 状态的 zone。OpenRouter 用于调用付费模型,必须至少充值 10 美元,才能把免费模型的额度提高到每天 1,000 次。Linode/Akamai 用于第 6 课,需要添加支付方式、启用 2FA,并创建一个有效期约 90 天的 token。Kaggle 免费提供两块 T4 GPU,每周 30 小时;需要验证手机号,并在 notebook 中安装包时必须打开互联网。标记为 Live account 的检查命令必须使用真实账号,预期结果已经写明。A domain before Session 2
The CDN lab puts s2.<root> behind Cloudflare's edge, so by Session 2 you need a
root domain you control and a Cloudflare zone that is Active. A domain is
a name you rent from a registrar for about $8–12/year; it is your asset, paid
by you, and it is not reimbursed. Register a root domain such as
example-course.net, not a subdomain: Cloudflare Free cannot host a subdomain as
its own zone.
Nameservers are the DNS records that tell the internet which servers answer for your domain. Changing them at the registrar to Cloudflare's two nameservers delegates the zone to Cloudflare; propagation can take up to 24 hours, so do this well before the session.
- Register the root domain at any registrar. Cloudflare Registrar is one option; a registrar you already use is fine.
- In Cloudflare choose Add a site, enter the root domain, and select the Free plan.
- At your registrar, replace the registrar's nameservers with the two Cloudflare nameservers shown in the dashboard.
- Wait, then prove the delegation.
macOS / Linux
dig NS <root-domain> +short
dig NS asks the public DNS which nameservers answer for the domain. Proof: two
names ending in .ns.cloudflare.com. An empty answer or your registrar's
nameservers means propagation is not finished; wait and run it again.
Windows (PowerShell)
nslookup -type=NS <root-domain>
nslookup performs the same query without dig, which Windows does not ship.
Proof: a non-authoritative answer listing the same two Cloudflare nameservers.
The second proof is the Cloudflare dashboard showing the zone as Active. Capture that screenshot; it is what Session 2's pre-flight reads. If the zone is not active in time, the session's fallback is the instructor's demo hostname, and you say so in your evidence note — the fallback changes the comparison, not the submission's shape.
สรุปภาษาไทย
ก่อนเซสชัน 2 ต้องมีโดเมนระดับรากที่คุณเป็นเจ้าของและโซน Cloudflare ที่ Active โดเมนมีค่าใช้จ่ายประมาณปีละ 8–12 ดอลลาร์ จ่ายเองและไม่ได้รับการคืนเงิน ห้ามใช้ซับโดเมนเพราะ Cloudflare แพลนฟรีไม่รองรับซับโดเมนเป็นโซนของตัวเอง ขั้นตอนคือจดทะเบียนโดเมนระดับราก เพิ่มเว็บไซต์ใน Cloudflare เลือกแพลนฟรี แล้วเปลี่ยนเนมเซิร์ฟเวอร์ที่ผู้ให้บริการจดทะเบียนเป็นสองชื่อของ Cloudflare ซึ่งอาจใช้เวลาถึง 24 ชั่วโมง ตรวจสอบด้วย `dig NS中文总结
在第 2 课之前,必须有一个你拥有的根域名和一个处于 Active 状态的 Cloudflare zone。域名每年大约花费 8–12 美元,需自行支付且不予报销。不要使用子域名,因为 Cloudflare 免费套餐不支持把子域名作为独立 zone。步骤是:注册根域名,在 Cloudflare 中添加站点并选择免费套餐,然后在注册商处把 nameserver 改为 Cloudflare 的两个 nameserver;传播可能需要长达 24 小时。在 macOS/Linux 上用 `dig NS <root-domain> +short` 检查,在 Windows 上用 `nslookup -type=NS` 检查;正确结果是两个以 `.ns.cloudflare.com` 结尾的名称,并且仪表板显示该 zone 为 Active。如果 zone 尚未就绪,就以教师的演示主机作为备用方案,并在证据记录中说明。SSH keys
An SSH key pair is two files: a private key that stays on your laptop and a
public key that you upload to the provider. The provider stores the public
half; when you connect, the two halves perform a challenge that proves you hold
the private half without ever sending it. That is why there is no password to
leak, and why losing the private key locks you out. ed25519 is the current
default: a short key that is hard to get wrong.
Generate the key once, in Session 0, and keep it for the course. Session 1
imports this same key into DigitalOcean; if its brief still shows the
ssh-keygen step, skip it when ~/.ssh/do_s1 already exists — regenerating would
overwrite the key. The checks below are identical either way.
macOS / Linux
ssh-keygen -t ed25519 -C "course-s0-<your-id>" -f ~/.ssh/do_s1
ssh-keygen creates the pair. -t ed25519 selects the algorithm, -C sets a
comment that identifies the key in provider listings, and -f writes the private
half to ~/.ssh/do_s1 and the public half to ~/.ssh/do_s1.pub. It asks for a
passphrase; use one, because it protects the private key if the laptop is lost.
Proof: it prints Your identification has been saved in /Users/<you>/.ssh/do_s1
and a SHA256: fingerprint, and ls ~/.ssh/do_s1* lists exactly two files.
chmod 600 ~/.ssh/do_s1
stat -f %Lp ~/.ssh/do_s1
chmod 600 makes the private key readable and writable only by you. stat -f %Lp
prints the permission bits in octal. Proof: 600. OpenSSH refuses to use a key
with looser permissions, so this check prevents the "bad permissions" error later.
On Linux use stat -c %a ~/.ssh/do_s1 for the same proof.
cat ~/.ssh/do_s1.pub
This prints the public key, which is safe to share. Proof: one line starting
ssh-ed25519 AAAA… and ending with the comment course-s0-<your-id>. Never print
the private key: the file without .pub is the one that must never leave the
machine.
Windows 11 (PowerShell)
ssh-keygen -t ed25519 -C "course-s0-<your-id>" -f "$HOME\.ssh\do_s1"
Windows OpenSSH provides the same ssh-keygen. The paths are
$HOME\.ssh\do_s1 (private) and $HOME\.ssh\do_s1.pub (public). Proof: the same
two-line success message and a SHA256: fingerprint. If $HOME\.ssh does not
exist, ssh-keygen creates it.
icacls "$HOME\.ssh\do_s1"
icacls prints the file's access control list. Proof: your account can read the
file and no other standard user can. SYSTEM and Administrators normally
remain and Windows OpenSSH accepts them. If another user appears, remove the
inherited entries and grant only yourself:
icacls "$HOME\.ssh\do_s1" /inheritance:r /grant:r "$($env:USERNAME):(R)"
/inheritance:r removes inherited entries and /grant:r replaces the grant with
read-only access for your account. Proof: Successfully processed 1 files and a
second icacls listing that contains only your user. Windows OpenSSH checks these
permissions and refuses the key when another account can read it.
Get-Content "$HOME\.ssh\do_s1.pub"
This prints the public key. Proof: the same ssh-ed25519 AAAA… line.
Upload the public key to DigitalOcean
Panel path: Settings → Security → SSH keys → Add SSH key; paste the contents
of do_s1.pub, name it course-s0-<your-id>, and save. A Droplet created with
this key accepts key-only login; DigitalOcean disables password authentication on
such Droplets.
After Session 1 creates your doctl token, the CLI path is:
doctl compute ssh-key import course-s0-<your-id> --public-key-file ~/.ssh/do_s1.pub
--public-key-file uploads only the public half; the private half never leaves
the laptop. Proof: doctl prints the new key's numeric ID and fingerprint.
doctl compute ssh-key list --format ID,FingerPrint,Name --no-header
ssh-keygen -E md5 -lf ~/.ssh/do_s1.pub
The first command lists the account's keys; the second prints the local key's MD5
fingerprint in the same form. Proof: the FingerPrint value in the doctl row
matches the MD5: value from ssh-keygen — that equality is the evidence that
the key you hold is the key the account holds.
Screenshot rule. Capture the fingerprint, never the private key. If you paste
the public key into the panel, the public key is safe; the file without .pub is
not.
สรุปภาษาไทย
กุญแจ SSH มีสองส่วน คือกุญแจส่วนตัวที่เก็บไว้บนเครื่องและกุญแจสาธารณะที่อัปโหลดให้ผู้ให้บริการ ผู้ให้บริการเก็บเฉพาะครึ่งสาธารณะ และการเชื่อมต่อพิสูจน์ว่าคุณถือครึ่งส่วนตัวโดยไม่ส่งมันออกไป จึงไม่มีรหัสผ่านให้รั่ว แต่ถ้าสูญเสียกุญแจส่วนตัวจะเข้าเครื่องไม่ได้ ed25519 เป็นค่าปริยายปัจจุบัน สร้างด้วย `ssh-keygen -t ed25519 -C "course-s0-中文总结
SSH 密钥分为两部分:保存在本机的私钥,以及上传给提供商的公钥。提供商只保存公钥那一半;连接时用挑战证明你持有私钥那一半,而不会把它发送出去,因此没有密码会泄漏;但如果丢失私钥,就无法再登录机器。ed25519 是当前默认值,用 `ssh-keygen -t ed25519 -C "course-s0-<your-id>" -f ~/.ssh/do_s1` 创建,并用 `chmod 600` 设置私钥文件权限。在 macOS 上用 `stat -f %Lp` 检查应得到 `600`,在 Linux 上用 `stat -c %a`。在 Windows 上使用 Windows OpenSSH 的 `ssh-keygen`,再用 `icacls` 检查权限,只保留你自己的账号。只通过 Settings → Security → SSH keys 页面或 `doctl compute ssh-key import` 上传 `.pub` 文件。证据是 `doctl compute ssh-key list` 中的 fingerprint 与 `ssh-keygen -E md5 -lf` 的输出一致。禁止截图或打印私钥。Safe-data policy
Every lab, screenshot and submission uses only the course's synthetic media pack: generated colour-bar clips, generated thumbnails, 20 EXIF-tagged photos of no real person, 100 transcripts with fabricated contact details, the caption catalogue and the Fashion-MNIST subset. The pack contains no real people, client footage or personal data. The rule is simple and absolute: never upload real personal media, or real account identifiers, emails or tokens in a screenshot, to any cloud service.
Why the rule is absolute. A data class is public, internal, personal
or sensitive; the class belongs to the asset and to the context. A colour-bar
still is internal on its own, but the same file with GPS in its EXIF and a caption
naming a student is sensitive. Redaction removes or replaces identifiers, but
it is a control, not anonymisation: a combination of benign details can re-identify
a person, and a model can invent an identifier that was never there. EXIF
metadata carries GPS, device model, capture time and a thumbnail; stripping it
before upload removes those fields while leaving the pixels unchanged. Retention
at a provider endpoint is a separate choice: the model id does not tell you the
endpoint's policy, and OpenRouter's in-region routing has no Asia option on the
student plan, so processing happens outside Thailand. The course therefore does
not rely on a cross-border transfer basis for real media; it sends synthetic data
only and says so in the governance rule.
The policy has four practical parts:
- Use only the supplied pack. Do not substitute your own photos, transcripts or client files, even if they look harmless.
- Strip before upload. The S8 pipeline removes names, emails, phones and student IDs, and strips EXIF, before any request leaves the controlled environment. The evidence is the metadata diff, not a promise.
- Keep credentials out of the kit. Tokens and private keys go in a shell
session or the provider console, never in a file that enters the repository.
The kit's
.gitignoreis a backstop, not permission. - Record the acknowledgement. Every lab brief carries the same safe-data acknowledgement sentence; you record it in Classroom, not in the kit.
After setup's step 4, this lists the pack:
ls assets/synthetic-media
Proof: the listing starts with client-clip-01.png, client-clip-01.metadata.json
and photo-01.jpg through photo-20.jpg, plus people.json and
manifest.json — all course-generated, none of it yours. If you see a file you
supplied, you are in the wrong folder or you copied personal data in; stop and
remove it before continuing.
สรุปภาษาไทย
กฎข้อมูลปลอดภัยคือทุกแล็บ ทุกภาพหน้าจอ และทุกงานที่ส่ง ใช้เฉพาะชุดสื่อสังเคราะห์ของรายวิชา ซึ่งประกอบด้วยคลิปแถบสี รูปย่อ ภาพถ่าย 20 ภาพที่ติด EXIF โดยไม่มีบุคคลจริง บทถอดเสียง 100 รายการที่ข้อมูลติดต่อถูกสร้างขึ้น แคตตาล็อกคำบรรยาย และชุด Fashion-MNIST ห้ามอัปโหลดสื่อส่วนบุคคลจริง ภาพลูกค้า หรือข้อมูลระบุตัวตน อีเมล และโทเคนในภาพหน้าจอ ขึ้นคลาวด์ใด ๆ เด็ดขาด ชั้นข้อมูลคือ public, internal, personal และ sensitive และชั้นข้อมูลขึ้นกับบริบทของไฟล์ การปกปิดข้อมูล (redaction) เป็นมาตรการควบคุม ไม่ใช่การทำให้ไม่ระบุตัวตน เพราะรายละเอียดที่ดูไม่มีพิษภัยรวมกันอาจระบุตัวบุคคลได้ และโมเดลอาจสร้างข้อมูลระบุตัวตนขึ้นมาเอง ต้องลบ EXIF เช่น GPS รุ่นอุปกรณ์ เวลา และภาพย่อ ก่อนอัปโหลด และเก็บโทเคนกับกุญแจส่วนตัวไว้นอกคลังโค้ด หลังติดตั้งคิตใช้ `ls assets/synthetic-media` ตรวจว่ามีเฉพาะไฟล์ของรายวิชา และบันทึกข้อตกลงข้อมูลปลอดภัยใน Classroom中文总结
安全数据规则是:每个实验、每张截图和每次提交都只能使用课程的合成媒体包,其中包括生成的色条片段、生成的缩略图、20 张带有 EXIF 且不含真实人物的照片、100 条联系方式为虚构的转录文本、字幕目录以及 Fashion-MNIST 子集。绝对禁止把真实个人媒体、客户影像,或截图中的身份信息、邮箱和 token 上传到任何云。数据分级为 public、internal、personal 和 sensitive,分级取决于文件的使用场景。脱敏(redaction)是一种控制措施,而不是匿名化,因为看似无害的细节组合起来仍可能识别出个人,模型也可能自行生成身份信息。上传前必须删除 EXIF,例如 GPS、设备型号、时间和缩略图,并把 token 和私钥放在代码仓库之外。安装 kit 后,用 `ls assets/synthetic-media` 检查其中只有课程文件,并在 Classroom 中记录安全数据确认。Assessment
The course is worth 100 points: 70 weekly and 30 capstone. Google Classroom is
the system of record for submissions, feedback and grades; this site carries
the instructions, not your marks. The join link is the owner-only placeholder
<Classroom join link>; your instructor fills it in before publication.
| Component | Points | What you submit |
|---|---|---|
| Weekly research | 40 (4 × 10 sessions) | An 800–1,200 word answer to the session's research question, with 5–8 sources |
| Weekly lab | 30 (3 × 10 sessions) | The evidence bundle: the scenario artifact, a 150–250 word "what this proves" note, and a teardown log |
| Capstone proposal | 5 | The design and feasibility case for the catalogue's AI-tagging phase |
| Capstone implementation | 15 | The working pipeline and the checkpointed batch results |
| Capstone report and presentation | 10 | The evaluation, governance rule, budget reconciliation and the talk |
| Total | 100 |
The three programme learning outcomes (PLOs) are weighted 30 / 40 / 30:
| PLO | Points | Where it is evidenced |
|---|---|---|
| PLO1 — evaluate architectures | 30 | Weekly research answers and the capstone proposal |
| PLO2 — build and operate | 40 | Weekly lab evidence and the capstone implementation |
| PLO3 — analyse responsibility and risk | 30 | Weekly research and lab evidence, the risk register, and the capstone report |
Each weekly activity uses a four-level rubric. Research is scored 4 / 3 / 2 /
1; a lab is scored 3 / 2 / 1 / 0. A zero is reserved for no submission,
unsafe data or academic misconduct. The weekly PLO.md file on each session page
shows that session's rows and descriptors, and the master reconciliation is
PLO.md.
The capstone reuses the campus media unit's 5,000-asset catalogue: add AI
tagging, with a governance rule, inside a fixed monthly budget, on the course
stack. The brief and the ten-item evidence bundle are in
samples/capstone/README.md. Proposal 5 points,
implementation 15, report and presentation 10.
What "evidence" means. A lab is not "I ran the commands". The bundle is the scenario artifact (the JSON, screenshots and files the brief names), a short note saying what the artifact proves and what it does not, and a teardown log showing the account returned to its starting state. A command output without the note is not evidence; a note without the artifact is not either.
สรุปภาษาไทย
รายวิชามี 100 คะแนน แบ่งเป็นคะแนนรายสัปดาห์ 70 คะแนน (วิจัย 40 + แล็บ 30) และแคปสโตน 30 คะแนน (ข้อเสนอ 5, การลงมือทำ 15, รายงานและการนำเสนอ 10) Google Classroom เป็นระบบบันทึกหลักสำหรับการส่งงาน ผลตอบกลับ และเกรด เว็บไซต์นี้มีคำสั่งและคำอธิบาย ไม่มีคะแนนของคุณ แต่ละสัปดาห์มีคำถามวิจัย 800–1,200 คำพร้อมแหล่งอ้างอิง 5–8 แหล่ง และแล็บที่ส่งเป็นชุดหลักฐานสามส่วน คือชิ้นงานสถานการณ์ บันทึกสั้น 150–250 คำว่าสิ่งนี้พิสูจน์อะไร และบันทึกการรื้อทรัพยากร PLO มีน้ำหนัก 30/40/30 คือ PLO1 ประเมินสถาปัตยกรรม 30, PLO2 สร้างและดำเนินงาน 40, PLO3 วิเคราะห์ความรับผิดชอบและความเสี่ยง 30 รูบริกสี่ระดับใช้กับงานวิจัย 4/3/2/1 และแล็บ 3/2/1/0 โดยศูนย์สงวนไว้สำหรับไม่ส่งงาน ใช้ข้อมูลไม่ปลอดภัย หรือทุจริตทางวิชาการ แคปสโตนใช้แคตตาล็อกสื่อของหน่วยมหาวิทยาลัยพร้อมกฎธรรมาภิบาลและงบประมาณรายเดือน中文总结
本课程共 100 分,分为每周 70 分(研究 40 + 实验 30)和 capstone 30 分(提案 5、实施 15、报告与展示 10)。Google Classroom 是提交、反馈和成绩的权威记录系统;本网站提供说明和解释,不显示你的成绩。每周有一个 800–1,200 词的研究问题,附 5–8 个来源,以及一个以三部分证据包提交的实验:情境产物、150–250 词说明它证明了什么的短文,以及资源拆除记录。PLO 权重为 30/40/30,即 PLO1 评估架构 30、PLO2 构建与运营 40、PLO3 分析责任与风险 30。四级评分标准用于研究 4/3/2/1 和实验 3/2/1/0,零分保留给未提交、使用不安全数据或学术不端的情况。capstone 使用大学媒体的媒体目录,并附带治理规则和每月预算。Readiness checklist
Work down this table before Session 1. Every row ends in something you can see —
a version line, a permission bit, a summary: line, a panel state — and you record
that result in Classroom under the Session 0 readiness check
(<Classroom join link>). A row you cannot prove is a row to fix before the first
lab.
| # | Item | Command or panel path | Result you record |
|---|---|---|---|
| 1 | uv installed | uv --version |
uv 0.13.x or newer |
| 2 | git installed | git --version |
git version 2.x |
| 3 | VS Code and Jupyter | code --version; code --list-extensions |
a 1.1xx.x line; the list contains ms-toolsai.jupyter and ms-python.python |
| 4 | doctl installed | doctl version |
a version line and API endpoint: https://api.digitalocean.com/ |
| 5 | Wrangler installed | wrangler --version |
4.x |
| 6 | linode-cli installed | linode-cli --version |
linode-cli v5.x |
| 7 | aws installed | aws --version |
aws-cli/2.x (or a 1.x line) |
| 8 | jq installed | jq --version |
jq-1.8.x |
| 9 | SSH key pair exists | ls ~/.ssh/do_s1* (PowerShell: dir "$HOME\.ssh\do_s1*") |
exactly two files: do_s1 and do_s1.pub |
| 10 | Private key is private | stat -f %Lp ~/.ssh/do_s1 (Linux stat -c %a; Windows icacls) |
600; on Windows, no other standard user can read it |
| 11 | Public key is correct | ssh-keygen -lf ~/.ssh/do_s1.pub |
one SHA256: fingerprint ending with course-s0-<your-id> |
| 12 | Kit download intact | shasum -a 256 -c SHA256SUMS (Linux sha256sum -c; Windows compare hashes) |
cloud-tech-kit.zip: OK, or two matching hashes |
| 13 | Environment built | uv run --no-sync python -V inside the kit |
Python 3.14.x |
| 14 | Data fetched | uv run --no-sync python fetch_data.py --all |
a summary: line with a non-zero file count |
| 15 | A notebook runs | VS Code Run All, or bash start-jupyter.sh then Run All Cells |
the last cell completes; outputs show the CSV rows read |
| 16 | DigitalOcean safe | Panel: Account → Settings, Teams → Members, Billing → Spend alerts | 2FA Enabled with a backup method; instructor listed as Resource Viewer; $10 alert present |
| 17 | Cloudflare ready | npx wrangler whoami after the S4 token; or the dashboard login before it |
the account table and token permissions; zone Active after the domain step |
| 18 | OpenRouter ready | curl -s https://openrouter.ai/api/v1/credits -H "Authorization: Bearer $OPENROUTER_API_KEY" | jq . |
JSON with data.total_credits and data.total_usage |
| 19 | Linode/Akamai ready | linode-cli account view --format company,balance,balance_uninvoiced --text --no-headers |
three columns, not Invalid Token |
| 20 | Kaggle ready | Kaggle notebook: !nvidia-smi; Settings → Accelerator |
a Tesla T4 line; GPU T4 ×2 selected |
| 21 | Safe data acknowledged | Read Safe data | you recorded the acknowledgement in Classroom |
Rows 16–20 are Live account checks: the expected result is given, and the instructor's sanitized run is in the course runbook (#112). If a provider is unavailable on the day, use the session's documented fallback chain and say which substitution you made in the evidence note — a fallback changes the comparison, not the submission's shape.