Jiajun Li — personal website

A static academic and engineering portfolio for GitHub Pages. Pages are served directly from the repository root, with no runtime dependencies or third-party scripts.

The site is laid out like a technical blog: one reading column, monospace headings, and notes near the top of the homepage. See Design system.

Local preview

python3 -m http.server 8000

Open http://localhost:8000. Stop with Ctrl+C.

Design system

assets/style.css defines its tokens on :root. Use them instead of literal values:

• Layout: one reading column, --page-width (760px). Sections stack; use .split or .section for new ones.
• Type: six sizes, --text-label to --text-display. --text-label (13px) is the working minimum; tests/test_site.py enforces 12px. Headings and site chrome use --mono, body text uses --sans.
• Spacing: multiples of 8px (--space-*).
• Links: content links are blue and underlined; titles and site chrome stay in ink.
• Code: inline code sits on a soft chip; code blocks use a dark panel (--code-bg).

Source Sans 3 (body, OFL-source-sans.txt) and IBM Plex Mono (headings and code, OFL.txt) are self-hosted in assets/fonts/ under the SIL Open Font License. Source Sans uses its Latin, Latin Extended, and Greek subsets; the Plex Mono files are subset to include arrows and subscripts, and follow it in --sans as the fallback for those characters. Chinese text uses the system CJK face.

The homepage leads with quantitative research, machine learning, and real-time systems. The career timeline in content/pages/index.html uses --from and --to percentages to map dates onto a 2013–2027 axis. The Publications page places one linked mark per record from data/publications.json on the same axis.

Content and maintenance

• index.html: introduction, profile links, the latest note, a career timeline, four selected-project categories, and five compact selected publications.
• about.html: a two-paragraph profile, professional background, education, and research themes.
• publications.html: searchable full bibliography, paper links, and individual BibTeX citations.
• projects.html: QLeaf, GardenPipe, OptIndex, RL Tools, QuDrip, and the collaborative NESSi 2.0 contribution.
• contact.html: email, then LinkedIn, Google Scholar, and GitHub profiles.
• data/publications.json: reviewed bibliographic metadata and access links.
• content/pages/*.html: editable English page content, including the profile.
• content/pages.yaml: page titles and descriptions.
• content/projects.yaml: project descriptions.
• scripts/build_site.py: shared page layout and static generation.
• assets/style.css: responsive layout and print styling.
• assets/publications.js: progressive search, type/year filters, and reset behavior.
• assets/publications.bib: downloadable UTF-8 bibliography (biblatex/biber recommended).

Edit the publication dataset or files under content/, then regenerate:

python3 scripts/build_site.py

Commit both the source edits and the generated HTML/BibTeX. GitHub Pages does not need to run the generator. All publication records and links remain available without JavaScript. Search matches title, authors, venue, year, DOI, arXiv ID, and earlier titles, ignoring case and diacritics.

Languages

English pages remain at their existing URLs; Simplified Chinese pages are under /zh/. The language switch keeps the current page. The supplied portrait remains in assets/me.jpeg but is not displayed on the homepage.

Translations use authored overrides first, then a current automatic draft, then English. Initial Chinese drafts are included, so ordinary builds work offline. See the translation guide for paragraph overrides, complete Chinese notes, review reports, and API updates. Generated pages use relative links so they also work when opened directly as HTML files. Preview the generated root index.html, not the source fragments in content/pages/.

python3 scripts/update_translations.py
.venv/bin/python -m unittest discover -s tests -v

Homepage selection

The selected flags in data/publications.json control the homepage. The selection includes NESSi 2.0, both first-author Physical Review Letters papers, Aaram Kim’s Physical Review Letters paper on vertex-based treatment of light–matter systems, and the Nature Communications paper. The complete bibliography remains unchanged.

Selections display newest year first, with citation count as the same-year tiebreaker. Citation counts use the Scholar snapshot retrieved on 19 October 2026; a subsequent refresh was rate-limited. Counts are recorded as editorial metadata and are not displayed on the site.

Sources and editorial decisions

Checked on 19 October 2026:

• Google Scholar: all 46 profile entries and individual citation metadata were retrieved. Four duplicate entries are consolidated into their journal articles; their earlier titles and source links are retained in aliases.
• LinkedIn: direct access was blocked. The publicly indexed Swiss profile confirms the quantitative research background and computational sciences certificate. The owner supplied the complete profile text in this session, including QRT and IMC roles, garden leave, PSI fellowship, FAU research role, and degrees at Buffalo and Beijing Normal University. Employment dates come from that text. Degree dates were not supplied; only the separately verified 2017 doctoral thesis date is shown. Private analytics and social activity from the pasted profile are not included.
• GitHub: project descriptions were checked against the five personal repositories' current READMEs and the NESSi repository. QLeaf’s latency summary refers to measured benchmark configurations; hardware, methodology, and detailed results are linked from its project entry.
• Eckstein group publication list: identifies the additional photon-mediated spin interaction paper, subsequently published as Physical Review B 109, 085149 (2024). The arXiv record and Crossref metadata confirm its publication and authors.
• arXiv records: verified titles, author lists, journal references, and identifiers. Scholar omissions of η, Sr₂IrO₄, and 1T-TaS₂ were repaired from publication metadata.
• Crossref publisher metadata: checked journal metadata and version-of-record open licenses. Confirmed open-access papers use journal PDFs; other papers use arXiv PDFs. The open-access IOP proceedings paper uses its Scholar-listed publisher PDF. Publisher bot restrictions can prevent automated PDF checks even when a link is valid.

The bibliography contains 28 journal/proceedings articles, 4 standalone preprints, 1 thesis, and 10 meeting abstracts, totaling 43 distinct records. Meeting abstracts have record links, not invented paper downloads. Preprints and published versions are not double-counted. No citation counts are displayed because they change frequently.

Validation

The bilingual build has automated checks for translation precedence, stale-source behavior, complete authored notes, protected benchmark numbers/code, local links, assets, and language metadata. Browser checks cover the portrait, Chinese layout, language switch, translated diagram, and publication search. The Google API request is tested with a mock; no live API credentials are configured.

During the original bibliography check, thirty PDF links returned PDF content; two APS PDF endpoints returned HTTP 403 and the IOP PDF endpoint returned a bot challenge. Those three publisher-provided links remain in place.

Publish

Push the files to main in LiJ21/LiJ21.github.io. Configure the repository's Settings → Pages to deploy from the main branch and / (root) directory. The canonical site address is https://jjli.info/.

.nojekyll keeps the deployment a plain static site. This edit does not commit, push, or alter hosting settings.

Notes

Notes are authored in content/posts/*.md and rendered to blog/<slug>/index.html. blog.html lists published notes newest first, and feed.xml provides an RSS feed. The explicit slug keeps the URL stable when the date or title changes.

Install the build dependencies once in a virtual environment:

python3 -m venv .venv
python3 -m pip --python .venv/bin/python install -r requirements.txt

If the system lacks ensurepip, create the environment with python3 -m venv --without-pip .venv instead, then use the same pip command above. The existing python3 scripts/build_site.py command automatically uses .venv when its dependencies are not available in the current interpreter.

Create an article with YAML metadata followed by Markdown:

---
title: "A new note"
subtitle: "An optional subtitle shown below the article title"
slug: a-new-note
date: 2026-10-19
summary: "A short description for the notes list and RSS feed."
tags: [Systems, Machine learning]
draft: true
---

Start the article here. Use level-two headings for sections.

Markdown supports tables, fenced code blocks, links, images, and footnotes. Place figures in assets/posts/<slug>/ and reference them with root-relative paths such as /assets/posts/a-new-note/figure.svg. Article slugs must be unique, lowercase words separated by hyphens. Dates use YYYY-MM-DD.

Set draft: false to include a note in the site and feed, then run:

.venv/bin/python scripts/build_site.py
python3 -m http.server 8000

Visit /blog.html to preview. Commit the Markdown source, generated pages, updated navigation, and feed together. No server-side database or JavaScript is required for notes. Generation removes previously generated article pages when notes are deleted or returned to draft. Drafts are omitted from rendered pages and RSS, but their Markdown remains visible if the repository is public.

The first note follows QLeaf’s latency ablation, framework structure, CPU workers, and GPU optimizations. It links to a fixed source revision and includes six SVG diagrams under assets/posts/qleaf-inference-latency/. To edit the diagrams, change scripts/draw_qleaf_diagrams.py and run python3 scripts/draw_qleaf_diagrams.py. The generated SVG files are served directly, without a diagram-rendering runtime.

Search engine sitemap

Every build regenerates sitemap.xml with the main pages and published notes, using https://jjli.info. Drafts and the 404 page are excluded. robots.txt allows crawling and points to the sitemap. After deploying, submit https://jjli.info/sitemap.xml in Google Search Console → Sitemaps.

Each published page includes an absolute canonical URL and Open Graph URL matching its sitemap entry. Page titles and descriptions are generated with the content; the homepage title identifies Jiajun Li and the site's focus. The 404 page has no canonical URL and remains excluded from the sitemap.

Hosted journal PDFs

Fifteen unmodified APS journal PDFs are stored in assets/papers/, with download provenance and checksums in data/hosted-pdfs.json. Copyright or CC BY notices appear with the corresponding citations. Nine other APS journal copies remain unavailable; their existing arXiv download links are retained. See the PDF inventory for the remaining papers.