Site for Experimental Intelligence, a research lab that trains and post-trains small open models and measures what each step changes, including what gets worse. Sister organisation: experimentalmachines.org, which shares its team and its homepage design.
Next.js (App Router, TypeScript) + Tailwind CSS v4. Static export, no API routes.
npm run devOne page, told in order: a violet hero with two results drawn as slope charts, a strip of the four projects, then one section each for OpenGrad, Tool-choice contrast, the 151M from-scratch run and Taglish sentiment, then the team and contact.
lib/content.ts: site details, the project strip, Tool-choice contrast, the 151M run's headline figures, Taglish sentiment and the team.lib/opengrad.ts: OpenGrad Study 001 and its quantized builds, transcribed from the OpenGrad repository's audited artifacts.lib/datasets.ts: the Canonical v2 training corpus, with each source's licence and row counts.lib/run.ts: the 151M run'smetrics.jsonl, downsampled, and its benchmarks after each stage.
app/globals.css holds the design tokens. The page has no animation: every chart is drawn in full on load.
The eye mark is traced to vector from the original logo: public/brand/ei-mark.svg (violet) and ei-mark-white.svg (for the violet header). Everything else is rendered from those: the icons as the white eye, level and zoomed onto the iris so the tips run off the edges, on a violet (#7b3fe4) tile, the link previews on white.
app/favicon.ico(16, 32, 48) andapp/icon.svg: browser tabs. The SVG and the 16px layer sit a little closer (1.75× the tile width, against 1.6× elsewhere) so the swirl stays thick at tab size.app/apple-icon.png(180) andpublic/brand/icon-{192,512,maskable-512}.png: home screens, listed inapp/manifest.webmanifest.app/opengraph-image.pngandapp/twitter-image.png(1200×630), with.alt.txtbeside each: link previews.public/brand/ei-mark.png(1200 wide): the logo in the page's Organization structured data.
Pushes to main deploy to GitHub Pages through .github/workflows/pages.yml.