CUNY AI Lab website

This repository contains the CUNY AI Lab's Astro static website, including its content collections, tools portal, model registry, and request-access flow.

Project structure

/
├── public/              # Static assets and robots.txt
├── src/
│   ├── components/      # Shared header and footer
│   ├── content/         # Markdown pages and blog posts
│   ├── data/            # Structured page data
│   ├── layouts/         # Shared Astro layouts
│   └── pages/           # Site routes
├── astro.config.mjs
├── amplify.yml
├── package.json
└── tsconfig.json

Commands

Run these commands from the repository root with the pinned Bun 1.4.2 runtime:

Command	Action
bun install --frozen-lockfile	Install the locked dependencies
bun run dev	Start the local development server at localhost:4321
bun run check	Audit dependencies, run Astro's static check, and build the site
bun run build	Build the production site to ./dist/
bun run preview	Preview the production build locally
bun run preview:cloudflare	Serve the build through local Cloudflare Workers at localhost:4321
bun run deploy:cloudflare	Deploy static assets to the Cloudflare preview

Hosting

The public site remains at https://ailab.gc.cuny.edu on AWS Amplify. Current main also deploys to https://cail-website.ailab-452.workers.dev for reviewing the Cloudflare hosting path. Both hosts serve the same Astro build; there is no custom Worker runtime, storage, or application binding.

Build before running either local preview command. The Cloudflare preview preserves canonical links to the live site and tools. Request-access identity and form submission are allowed only from the canonical website origin; use the production application page for that flow.

A production move is separate: establish Cloudflare zone and domain routing, coordinate the CUNY DNS change, verify the canonical site, then retire Amplify. No email migration is planned.

CAIL Sandbox Docs

/sandbox-docs/ and its article routes are built from src/content/sandbox-docs/*.md. The initial import preserved all 14 published Markdown files from CUNY-AI-Lab/sandbox-docs at revision 74ce09d2bb6cc58b37e1f7296a3a01e988009222, including examples, tables, and expandable sections. src/data/sandbox-docs-source.json records that import's source revision and SHA-256 hashes. It is provenance for the initial import, not a requirement to keep future documentation frozen.

Make future documentation edits here through the normal website pull request and deployment workflow. Builds do not fetch GitHub Pages or depend on its runtime Markdown renderer. The original GitHub Pages site redirects to these routes, including legacy article hashes and direct article paths. Its retained Markdown files are a historical source snapshot, not the editing destination.

The docs use the shared website header, footer, fonts, and canonical URLs, with the original three sidebar groups plus links to the three supplementary articles. Relative .md links become website routes during rendering, and legacy links such as /sandbox-docs/#student-onboarding resolve to the article. Each article stores its title and original heading ID in frontmatter for the compact shared page hero; the landing introduction also lives in frontmatter. The landing page uses the title CAIL Sandbox Docs and a short introduction to its teaching, research, privacy, and experimentation focus. Student Onboarding now foregrounds the instructor course-access checkpoint and the course invitation, group, and channel enrollment process. Getting Started uses the current lab support address, ailab@gc.cuny.edu. Other imported article text is unchanged except for standardizing the name to CAIL Sandbox.

Live model registry

The registry loads the public Gateway /v1/catalog without credentials on page load and Refresh catalog. Each canonical catalog row renders once, with one exact API ID and copy control. IDs preserve the complete model variant, such as gpt-oss-20b. Gateway owns identity and routing; the website does not group, strip prefixes, display alternate IDs, or name the provider serving a model. Featured models render as separate cards above the full registry in Gateway's numeric recommendation order and also appear in the canonical rows below. Gateway owns recommended, tier, and order; the website has no static Featured list. Featured status changes prominence, not availability. Search and capability filters operate on the full registry. Malformed recommendation metadata makes a model nonfeatured without removing its valid canonical row.

Capabilities reflect Gateway's eligible-route union. A request still requires a route supporting its complete capability combination. Context and comparative prices describe the preferred offering, so they are not guarantees for every request. Models with similar names remain separate rows.

Gateway supplies clean model_name labels and optional sourced specifications with their own check date. Cards summarize size, weights, and license; the native Details and sources disclosure contains architecture and per-field evidence. The website has no static metadata lookup. Published size, architecture, weights, and license facts link to their sources. Checkpoint parameter counts are distinguished from advertised sizes. Missing facts are omitted. Malformed optional fields are ignored without removing valid models. An invalid specifications check date discards those specifications. Capability icons and the 100K long-context indicator come from the canonical model. New catalog models appear automatically, including non-text routes.

The check time applies to catalog availability, route capabilities, context, and prices. A malformed row is omitted without hiding valid rows; malformed optional pricing stays unpublished. A malformed envelope or a response with no usable, unambiguous rows clears previous results and offers retry. An empty catalog and a search with no matches have separate messages. JavaScript is required for the interactive list; a public JSON catalog link remains usable without it. No inference request is sent by browsing, filtering, or copying IDs.

Prices come only from Gateway's optional pricing field, in USD per million input/output tokens. Missing or malformed prices stay unknown; zero is displayed only when the catalog reports zero. Gateway labels a price as a standard rate or as from when it is a starting rate. The selected route can cost more, and additional fees may fall outside these token rates. The page presents prices for the model without inferring or naming a provider. They are not personal spending, quota balances, or estimates of a particular request's cost.

The website is the caller and the existing public Gateway catalog is the receiver; Gateway must supply sourced metadata for the reference facts to appear. Browser tests substitute catalogs for deterministic filtering, failure, and identity cases. Rollout verification also loads the deployed page against the real public Gateway without intercepting that request, compares its usable IDs and recommendation order with that response, and does not freeze an expected model count. No API key or paid inference is needed. Gateway does not currently publish snapshot or completeness metadata, so this check cannot prove that the returned response is a complete registry snapshot.

Accessibility checks

After building, run bun run preview:cloudflare, then uv run --with playwright python tests/accessibility_browser.py. CI runs this alongside the keyboard and form journeys. Axe checks the rendered pages and expanded tool/model states against WCAG A/AA rules, with representative mobile views. It does not certify conformance or replace testing with assistive technology. The scanner is a development dependency, not a website script.