Talks: one page per talk, generated from TOML records

[gvegayon]

Talks only existed in the Google Calendar: the site linked out to calendar entries, and nothing about a past talk (abstract, speaker bio, slides, recording) survived anywhere. This adds a small pipeline so every talk gets a permanent page on the site.

How it works

1. A talk is a TOML file in _data/talks/, e.g. 2026-09-04-george-vega-yon.toml — title, date/time, location, Zoom, slides, recording, abstract, and one [[speakers]] block per speaker (name, affiliation, website, photo, bio).
2. make talks (scripts/generate_talks.py) turns each record into _talks/<name>.md, published at /talks/<name>/.
3. Both files get committed. .github/workflows/talks.yml re-runs the generator with --check on every PR, so the pages can never drift from the records.

Why generate instead of reading TOML at build time? GitHub Pages builds Jekyll in safe mode, so custom plugins (the usual way to read TOML) are unavailable — the pages have to exist in the repo.

What is in this PR

_data/talks/*.toml	178 talk records, 2020 through 2026
_data/talks/_TEMPLATE.toml	copy-me starting point, with every field documented
scripts/generate_talks.py	TOML → _talks/*.md; validates records, removes orphans, --check mode for CI
scripts/import_calendar_talks.py	seeds records from the calendar's public iCal feed
_layouts/talk.html	the talk page: abstract, speaker bios and links, slides/recording buttons
talks.md	/talks/ — upcoming talks, then past talks grouped by year
_includes/next_talks.html	now reads the talk records instead of fetching Google Calendar in JavaScript
seminar.md	next talk + next five upcoming talks + link to the archive; the calendar embed stays for subscribing
_config.yml	talks collection, and future: true (Jekyll hides future-dated documents otherwise, which would have hidden every upcoming talk)
README.md	how to add and update a talk

Verified by building the site with Jekyll 3.10 (the GitHub Pages version): 178 talk pages + the archive, no warnings.

About the seeded records

The 178 records were parsed from the calendar's iCal feed. Calendar descriptions are free-form, so this is a first pass: every imported record carries needs_review = true under [meta]. What came through well: titles, speaker names, dates, locations, and most abstracts and bios. What needs a human pass:

• 4 talks whose calendar entry never had a real title (they read TBA, Orientation, VLM Agents)
• ~78 older entries with no abstract in the calendar, and ~99 with no bio
• affiliations are sparse for the 2020–2022 entries
• no slides or recordings — those links were never in the calendar and have to be filled in by hand

Re-running make import-talks never overwrites an existing file, so corrections are safe.

Two things worth a decision:

• Zoom links on past talks. The records keep them, but the page only shows the Zoom link while a talk is still upcoming — publishing ~150 old join links (several with embedded passcodes) seemed worse than useless. Easy to change in _layouts/talk.html if you disagree.
• Scope of the archive. I imported everything back to 2020. If you would rather launch with, say, 2024 onward, deleting the older TOML files and re-running make talks is all it takes.

🤖 Generated with Claude Code

[github-actions]

Thank you for your contribution @gvegayon 🚀! Your jekyll-site is ready for download 👉 here 👈!
_{(The artifact expires on 2026-11-17T01:19:27Z. You can re-generate it by re-running the workflow here.)}

[gvegayon]

Pushed the follow-ups from review:

Speaker websites. The talk page now shows an explicit link (engineering.nyu.edu) under the speaker's name and affiliation, in addition to the name itself being a link. 91 of the imported records have a website; the rest are blank in the calendar and need filling in by hand.

Tags. Every record now carries tags drawn from a fixed vocabulary of 23 topics. Distribution across the 176 talks:

machine learning 52	algorithms & theory 37	natural language processing 30	fairness & ethics 26
statistics 22	networks & graphs 20	deep learning 19	society & policy 19
large language models 18	visualization 16	biology & genomics 14	health & medicine 14
data management 13	education 13	optimization 13	human-centered computing 12
computer vision 10	climate & environment 8	robotics 8	physics & astronomy 7
causal inference 5	geospatial 5	privacy & security 4

scripts/tag_talks.py assigns them by matching title (weighted x3), abstract, and bio against the vocabulary, capped at 4 tags per talk. It skips any record that already has tags, so hand-picked tags always win — the automatic pass is a starting point, not a permanent owner. 10 records with no abstract or bio matched nothing and were left untagged. The vocabulary lives at the top of that script; adding or renaming a topic is a one-line edit followed by --overwrite.

Search. /talks/ now has a search box plus tag, speaker, and year filters. It all runs in the browser over the list the page already renders — no search service, no index to build, no dependency. Search covers title, speaker, affiliation, tags, location, date, and the full abstract text. Filters compose (tag + year + text), empty year groups collapse, and the result count updates as you type. Filter state lives in the URL, so /talks/?tag=robotics or /talks/?speaker=Anna%20Fariha is shareable — and the tags on each talk page link back into the filtered archive. Without JavaScript the complete list still renders; only the filter controls are hidden.

Verified against a local build: 176 talk pages, searching "robot" narrows to 9, the robotics tag to 8, "measles" finds the one talk that mentions it only in its abstract, and the speaker dropdown carries 169 names.

One thing worth flagging: wiring up the tags surfaced parsing bugs in the calendar import that I had missed on the first pass, so this push also re-imports the records. All of fall 2023 had Data Science Lecture Series. Speaker as the speaker name (that summary shape wasn't handled), some records had taken Zoom link: https://... or Talks will be held in LNCO 1100 as their title, and "Madison Golden and Kaylee Alexander" had been read as a title rather than two speakers. Those are fixed, non-talks (orientations, the Sandia info session) are now skipped, and the count moved from 178 to 176 records. A talk with no known title is now titled "Talk by <speaker>" rather than "TBA" — 5 of those.