Inference Engineering — Distilled

A bilingual field guide to Inference Engineering: the path from product constraints and model mechanics through hardware, runtimes, optimization techniques, modalities, and production operation.

The site contains 24 plain HTML pages: two catalogues plus eleven units in English and Traditional Chinese. Chapters 0–7 are designed as focused 8–12 minute reads.

View

Published site — https://datasystem.school/inference-engineering-distilled/

• English catalogue — index.html
• 繁體中文目錄 — zh/index.html
• English unit — chapters/<slug>.html
• 繁體中文單元 — zh/chapters/<slug>.html

The URL owns the language. Switching languages preserves the current route, query string, and section anchor.

A Traditional-Chinese browser that lands on an English page is sent to the matching zh/ page before first paint, keeping the query string and anchor. The redirect runs one way only — a zh/ page never bounces anyone — so a Chinese link stays readable for whoever you send it to. Choosing a language from the switch records it in dss_lang and is honored from then on in both directions, shared with the other datasystem.school sites. Simplified locales stay on English, since the site ships Traditional Chinese only.

What's here

Path	Purpose
tools/build/book.json	Ordered bilingual book registry and reading times
tools/build/chapters/*.json	Source-of-truth bilingual content, claim references, and semantic diagrams
tools/build/reviews/*.json	Private content/render review attestations
tools/build/generate.py	Deterministic 24-page generator
tools/build/validate.py	Content, originality, rendering, and publication gates
tools/qa/	Exact 72-state responsive WCAG 2 AA harness
index.html, chapters/, zh/	Generated static site
styles.css	Shared low-contrast Neo-Brutalist Blueprint

Never edit generated HTML. Change tools/build/chapters/*.json, the registry, renderer, or shared stylesheet, then regenerate and verify the published files.

Generate and validate

Validate the bilingual source model:

python3 tools/build/validate.py content

Generate the site:

python3 tools/build/generate.py --output-root .

Check that committed files exactly match a fresh build:

python3 tools/build/generate.py --output-root . --check

siteOrigin in tools/build/book.json holds the authorized deployment base, so both commands emit canonical URLs, hreflang URLs, and sitemap.xml against it. The committed tree is therefore byte-identical to what Pages publishes, and the workflow fails if book.json and the live Pages base URL ever disagree. Change the deployment base in book.json, never by editing generated HTML.

An origin-less build omits canonical URLs, hreflang URLs, and sitemap.xml — useful for previewing from an arbitrary location, though it will not match the committed tree:

python3 tools/build/generate.py --output-root /tmp/preview --no-site-origin

Local preview

python3 -m http.server 9999 --bind 127.0.0.1 --directory .

Then open http://127.0.0.1:9999/.

Publish to GitHub Pages

.github/workflows/pages.yml publishes https://datasystem.school/inference-engineering-distilled/ on every push to main. Pages is configured with Settings → Pages → Source → GitHub Actions; no branch or /docs folder is used.

The workflow rebuilds the site against the deployment base reported by actions/configure-pages, so canonical URLs, hreflang URLs, and sitemap.xml resolve correctly for a project site, a user site, or a custom domain without editing content.

Only site content is uploaded. The published artifact is exactly:

index.html  zh/  chapters/  styles.css  favicon.svg  sitemap.xml

The source PDF, tools/, tests/, docs/, tmp/, and the generator's build metadata are never staged, and the workflow fails if any unexpected entry reaches the artifact. Repository visibility is separate: in a public repo those files stay browsable on GitHub itself even though Pages does not serve them.

Source and language policy

Original source — Inference Engineering by Philip Kiely, published by Baseten: https://www.baseten.co/inference-engineering/

The original PDF is the authority. This companion follows every approved printed-page range, points claim clusters back to source ranges, and uses independently composed HTML/CSS diagrams. Its English and idiomatic Traditional Chinese prose paraphrase the book; it does not reproduce the original figures or long passages. Original book text and figures remain the copyright of their authors and publisher.

Hardware, vendor, and software details are labeled as the book's January 2026 edition snapshot. Proper nouns and established technical acronyms stay in English where that is standard Taiwan engineering usage.

License

MIT © 2026 Data System School.

The license covers this repository's own work: the generator, validators, QA harness, tests, the bilingual prose, and the independently composed HTML/CSS diagrams. It does not and cannot grant rights to Inference Engineering itself — the original text and figures remain the copyright of their authors and publisher, as described above.

Full verification

Run the complete source, reading-time, review-evidence, originality, determinism, publication, and site audit:

python3 -m unittest discover -s tests
python3 tools/build/validate.py all --pdf "Inference Engineering.pdf" --root .

The full browser evidence is produced separately against a local preview with the pinned 72-state runner:

python3 tools/qa/run_accessibility.py \
  --base-url http://127.0.0.1:9999/ \
  --output-root tmp/qa