You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Landing: six curated showcases on the gallery stage, with calmer autoplay - #499
The landing page's gallery stage gets a full curation pass: which apps appear, in what order, and how the stage behaves while you watch it.
What's on the stage now — six showcases, opening with the two new AI-assisted apps:
Which US counties are alike? (new) — k-means on 3,152 counties; click one and its closest matches light up nationwide
Lending Club loans, in plain English (new) — querychat filters 10k loans; lasso the scatter and Claude describes exactly those loans
Ships slow down for whales — 2.68M AIS pings on a GPU map
Every star within 330 light-years — 256k Gaia stars linked across three views (sky tiles now load from IRSA; CDS's 404)
The Shiny issue atlas — promoted from runnable-only to a full gallery page
Live trading desk — restyled
How the stage behaves — a pass of autoplay fixes, each for something visibly wrong:
The stage only starts animating once scrolled into view (it used to be mid-cycle by the time you got there)
Hover/focus pauses in place and resumes where it left off, instead of jumping to the end
Clicking a showcase plays it from the start, even with the pointer on the stage
Only the screenshot pauses the stage — the app list doesn't — so a clicked app actually plays
The stage check in check_site.py now walks a visitor's whole session with these behaviors. Also included: a rewritten lede ("Ship delightfully rich experiences … directly from Python") and screenshots re-shot at 1440×900 so nothing important falls below the fold.
Verification
Run make docs && make docs-preview, open http://localhost:8780/, and scroll to the gallery stage: it starts animating only when scrolled into view, cycles counties → loans → whales → stars → issues → trade, pauses in place on hover, and plays a clicked showcase from the start. make docs-check exercises all of this headlessly (see the stage check in website/_tools/check_site.py).
Promotes the issue atlas from a runnable-only example to a full gallery
page (index.md folded in from README.md), fixes an #asked_twice
display:contents bug that let the twin list overflow unclamped, and
regenerates the trading-desk landing/card crops now that Task 1 replaced
their source screenshots.
Re-shoots whale-slowdown's hero as the true unfiltered state and adds a
California region + speed brush after-frame; adds a no-selection before
frame for nearby-stars to pair with its existing Hyades hero. Adds both
apps' GALLERY_CROPS entries, fixing make_landing_images.py --check.
…stage
Replaces the landing page's gallery stage with the four most polished
showcases (big-data geo, science, AI, and a live business dashboard),
dropping the older, plainer-looking apps. Also corrects the gallery
listing's item count in check_site.py, which had drifted out of sync
with the page's real content before this branch.
The sidebar said clicking a point opens its GitHub issue; the app never
wires up that gesture (the atlas isn't given a url column, and a plain
click doesn't change the selection), so it was never true. Describes the
actual drag-to-select gesture instead, and re-shoots the two screenshots
that showed the old wording.
Corrects issue-atlas's documented accessor (the app reads the
EmbeddingAtlasWidget's own _predicate/selection() directly, not a
render_embedding_atlas convenience method that doesn't exist), the
stale line count after the tip-text fix, and restores alphabetical
order in make_landing_images.py's GALLERY_CROPS.
…440x900
CDS's own DSS2 color tiles 404 and the alaskybis mirror times out, so the
stars app's sky panel never loaded. The whales 'Didn't slow down' table
fell below the fold at 1280x800.
The stage's progress bar and gesture animation ran from page load, so by
the time a visitor scrolled down to it, it was mid-cycle. It now waits
until the screen is half in view, stops when scrolled away, and restarts
the current showcase on return.
Hovering used to jump the progress bar to the end and leave the image
mid-gesture, then restart the app on leaving. It now freezes the bar,
crossfade, zoom and caption where they are, and continues from there.
…tage
Clicking a showcase while hovering the stage jumped straight to its
final frame with a full progress bar. It now plays from the start. Only
keyboard focus pauses the stage, so the focus a click leaves on a button
no longer keeps it paused after the pointer moves away.
…ans) showcases
loan-chat: querychat filters and queries 10,000 Lending Club loans in plain
English; lasso a rate band on the plotly scatter and a registered chatlas tool
lets Claude describe exactly those loans (answers checked against pandas).
county-clusters: k-means on 3,152 US counties (County Health Rankings 2025, no
race features); click the silhouette chart to pick k, click a county to see its
five closest matches nationwide drawn on the map.
Both join the landing stage (4 -> 6 showcases) with cards and landing images.
Clicking an app in the list happens with the pointer on the stage, so
after a click the stage either jumped to the app's end (paused) or
stopped pausing on hover. Hovering or keyboard-focusing the screenshot
now pauses; the app list doesn't, so a picked app plays.
Adds a 'stage' site check that walks a visitor's whole session with the
stage: before/after scrolling into view, pause and resume timing, list
hover, clicks, keyboard focus, scrolling away and back.
- loan-chat: opt the scatter into the selection accessor
(listen_to=["selected"]) so the lasso header and chatlas tool work
- check_site: fix the keyboard-stage assertion the showcase reorder
missed (3rd item is now the whales showcase)
- landing: gate stage autoplay on intersectionRatio >= 0.5, not
isIntersecting, so scrolling away through the threshold stops the
stage instead of restarting it
- county-clusters: derive the silhouette series from the cached
Clustering fits instead of fitting all nine k-means models twice
- landing: reset the pause flag/class in go() so a keyboard-picked
showcase can't leave the progress bar frozen against live timers
- loan-chat: carry grade in customdata so the hover shows the letter
grade, not its hex color
- loan-chat: put the legend dummy swatches in each grade's legendgroup
so a legend click toggles the actual loans
- plotly guide: drop listen_to from the bars decorator in the
linked-highlight snippet so it matches its example file again (this
was failing check_docs.py and the deploy)
The reason will be displayed to describe this comment to others. Learn more.
🔵 Needs a closer look
The county data mixes retired Connecticut counties with replacement planning regions, double-counting overlapping geographies, while outdated boundaries omit the new regions and can trigger a KeyError in map updates.
No open issues 🎉
2 previously missed issues
In code that hasn't changed since last review
Add polygons and centroids for Connecticut planning regions
The committed parquet includes Connecticut planning-region FIPS 09110–09190, but this GeoJSON was last updated in 2019, before those county equivalents were introduced in 2022. Those records have no polygon or centroid: they cannot appear on the map, and a nearest-neighbor result containing one causes a KeyError in figures.update_map(). Please align the boundary vintage with the county data and verify that every retained FIPS has a polygon and centroid.
Exclude overlapping retired Connecticut counties from the dataset
Filtering only on nonzero county FIPS retains Connecticut's eight retired counties (09001–09015) alongside their nine replacement planning regions (09110–09190); both sets are present in the committed parquet. These overlapping geographies are then treated as independent counties during imputation, standardization, clustering, and profile calculations, effectively representing Connecticut twice. Please retain a single county-equivalent vintage before imputation and regenerate counties.parquet so the model compares nonoverlapping units.
CHR 2025 reports Connecticut only as the nine planning regions that
replaced its counties in 2022; the eight retired counties (09001-09015)
are entirely empty in the source, so they were pure imputation and
double-counted the state. Worse, plotly's counties GeoJSON predates
those FIPS (plus three newer Alaska county-equivalents and Oglala
Lakota County), so a nearest-neighbor hit on one of them raised a
KeyError in update_map().
prep_data.py now drops the retired counties (3,152 -> 3,144 rows) and
commits a normalized TIGER supplement for the 13 affected FIPS (the
Census endpoint sends no CORS headers, so the app can't fetch it at
runtime); app.py merges it over plotly's file. Every FIPS in the data
now has a polygon and centroid; app smoke-tested in headless Chromium.
Both "previously missed" findings were real, and the second one turned out to be the tip of it — fixed in 59484be:
CT planning regions: verified against the raw CHR 2025 file that the eight retired counties (09001–09015) are entirely empty — their parquet values were pure imputation, while the nine planning regions hold the real data. So rather than dropping the regions, prep_data.py now drops the retired counties (3,152 → 3,144 rows) before imputation and the parquet is regenerated.
Missing boundaries: the same coverage check surfaced four more data FIPS with no polygon in plotly's 2013-vintage file (three newer Alaska county-equivalents, Oglala Lakota County). Since the Census TIGERweb endpoint sends no CORS headers (can't be fetched at runtime in the browser), prep_data.py commits a normalized supplement — 13 current-vintage shapes with ids and ~1 m coordinates — which the app merges over plotly's file. Every FIPS in the data now has a polygon and a centroid, so the update_map() KeyError path is gone.
Verified end-to-end: all 3,144 FIPS have geometry, twins of a CT planning region and of Oglala Lakota resolve, and the app smoke-tested clean in headless Chromium.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
The landing page's gallery stage gets a full curation pass: which apps appear, in what order, and how the stage behaves while you watch it.
What's on the stage now — six showcases, opening with the two new AI-assisted apps:
How the stage behaves — a pass of autoplay fixes, each for something visibly wrong:
The
stagecheck incheck_site.pynow walks a visitor's whole session with these behaviors. Also included: a rewritten lede ("Ship delightfully rich experiences … directly from Python") and screenshots re-shot at 1440×900 so nothing important falls below the fold.Verification
Run
make docs && make docs-preview, open http://localhost:8780/, and scroll to the gallery stage: it starts animating only when scrolled into view, cycles counties → loans → whales → stars → issues → trade, pauses in place on hover, and plays a clicked showcase from the start.make docs-checkexercises all of this headlessly (see thestagecheck inwebsite/_tools/check_site.py).