Screenly Florida Silver Alert App

A full-screen Silver Alert board for digital signage in Florida. Each time the page loads it shows one person from the Florida Department of Law Enforcement's current Silver Alerts: their photo, their age, when and where they went missing, and a QR code for the official FDLE flyer. Alerts issued in the last 30 days come first. 911 and FDLE's 1-888-FL-MISSING line are always on screen.

Silver Alerts are for adults who go missing with an irreversible deterioration of intellectual faculties, such as dementia, often while driving.

Live: https://silver-alert.srly.io

Part of the Screenly signage family alongside missing-children, which it is modelled on. Like it, the board is a static site on GitHub Pages, built on @screenly-labs/signage-kit.

Where the data comes from

FDLE's Missing Endangered Persons Information Clearinghouse lists every current Silver Alert at:

https://mepic.fdle.state.fl.us/currentSilverAlerts.asp

That page is HTML only (no JSON or RSS) and it doesn't answer CORS requests, so the browser can't read it directly. Instead an hourly GitHub Actions job scrapes it, validates every field, and deploys the result to Pages next to the board:

/data/alerts.json        the alerts
/data/photos/{id}.jpg    one photo per alert (some are .png)

The browser only ever reads those two paths on its own origin. The QR code links to FDLE's own flyer for the alert, https://mepic.fdle.state.fl.us/Flyer.asp?ID={id}.

Safeguards

Parsing someone else's HTML is fragile, and showing the wrong person, or someone who has already been found, is worse than showing nobody. So the scraper refuses to guess, and the board refuses to trust anything stale.

Scraper (scraper/)

• One fixed HTTPS URL. 30 s timeout, 4 attempts with backoff, redirects refused, 2 MB size cap, text/html required.

• Parsed with parse5, the spec-compliant HTML parser, not regexes.

• The whole page is refused unless the heading reads "Current Silver Alerts", there is exactly one results table, and its header has exactly the eight expected columns in order. "Nobody is missing" is only accepted when FDLE's own "No Silver Alerts are active at this time." notice is present.

• Each row is checked on its own:

  • exactly 8 cells, each holding only the tags FDLE uses there (a stray <script>, <img> or link rejects the row);
  • the photo and name link to the same numeric flyer id;
  • names must be "Last, First" in letters only;
  • dates must be real dates;
  • FDLE's printed "Age" and "Age Missing" must match the ages computed from the dates, which catches shifted columns;
  • the person must have been 18 or over when they went missing;
  • the missing date can't be in the future or before 2008;
  • the place must be "City,ST".

  Bad rows are dropped with a logged reason, never patched up.

• If more than a quarter of rows fail, or all of them do, the run fails.

• Photo and flyer URLs are rebuilt from validated numeric ids, never copied from the page. Photos must be real JPEG or PNG files by signature (FDLE labels PNGs as image/jpeg), 1 KB to 2 MB. A bad photo falls back to initials.

• The finished dataset must pass the same validator the browser runs, with nothing dropped, before it is written.

• Any failure fails the job before the deploy step, so nothing is published and the previous deploy stays up.

Board (assets/static/js/)

• Revalidates every alert (parseDataset) and drops unknown fields.
• Refuses the whole file if it's the wrong schema, has over 300 alerts, or was generated more than 4 hours ago. Freshness is judged against the server's Date header, so a player with a wrong clock can't revive stale data. A refused file shows the standing appeal (911 + FDLE line) instead of a person.
• Everything is rendered with textContent. The only markup inserted is the locally generated QR SVG.
• There is no bundled or cached list.

If the hourly job stops (FDLE down for hours, the page layout changes, or GitHub disables the schedule after 60 days without repository activity), screens fall back to the appeal within 4 hours rather than showing old alerts. A failed run shows up as a red run in the Actions tab.

Stack

• Bun for package management, bundling, tests and the scraper (no npm/npx)
• TypeScript, strict
• Tailwind CSS v4, CSS-first config, compiled by the Tailwind CLI
• Biome for lint + format
• Self-hosted Bricolage Grotesque + Hanken Grotesk, vendored via signage-kit

Develop

bun install
bun run dev        # build, scrape live into dist/data, serve dist/
bun run build      # build the static site into dist/
bun run scrape     # scrape FDLE into dist/data (after build)
bun test           # parser, validator and manifest tests
bun run typecheck  # tsc --noEmit
bun run lint       # Biome (lint:fix / format to auto-fix)

Tests run against fictional fixtures (test/fixture.ts) that copy FDLE's markup exactly. Never commit real alert data or screenshots of it.

Supported resolutions

Verified landscape and portrait at 3840×2160, 1920×1080, 1280×720 and 800×480 (Raspberry Pi Touch Display).

Deploy

.github/workflows/deploy-pages.yml builds, tests, scrapes and publishes dist/ to GitHub Pages:

• on a CalVer tag (YYYY.N.P), to ship new code;
• hourly (:17), re-deploying the latest tag with fresh data;
• by hand (workflow_dispatch).

Pushing to master on its own does not deploy. CI (ci.yml) typechecks, lints, tests and builds every PR; it never scrapes.

One-time setup follows the app-store repo's docs/new-app-setup.md (Pages with build_type=workflow, the tag deployment policy, DNS silver-alert.srly.io → screenly-labs.github.io, then the custom domain).

Attribution

Alert data and photographs are published by the Florida Department of Law Enforcement and are shown unaltered apart from name casing, with a link to each official flyer and a standing credit on every screen. This app is not affiliated with or endorsed by FDLE.

License

AGPL-3.0-only (see LICENSE).