Skip to content
Augusta BhardwajSay hello

The Stack Behind 4ugusta.dev: Next.js 16, Zdog and Plain CSS

· 6 min read

Just to be clear, when I built this site, this was version 1. It was using 15k particles. Built with Three.js and scroll animations, GSAP. It looked very good in a screen recording. 150kb gzipped. And a visitor had not seen a single project yet. Rebuilt it in October 2026. The rule. The work is the hero.

All of the moving things on this page need to be as valuable as their respective bytes of data. Here's what it's running on now and the reason why.

Next.js 16, almost all static

Next.js 16 with the app router, and React 19. Nearly every page is a server component, all built time rendering. Only use client components when something needs to move, and respond to the hand (such as the illustrations on the hero section, the tilt on the projects tiles, and the demo terminals).

Fonts

The site uses self-hosted fonts with next/font/local. Using Bricolage Grotesque for the display fonts, and Hanken Grotesk for the body text. There was no need to ask the font CDN to do their thing for me, and no layout shift when switching to those fonts.

export const bricolage = localFont({
  src: "../../public/fonts/BricolageGrotesque-Variable.woff2",
  variable: "--font-bricolage",
  display: "swap",
});

Illustration in the hero: Zdog machine, not a particle cloud

In the illustration on the top of my home page, there's a machine with five tools orbiting around it. This machine is drawn with Zdog. Zdog is a 2d-ish pseudo-3d rendering library that draws 2d rounded shapes into a canvas element. You can drag the machine around, and it'll rotate. If you release the mouse, it'll continue to rotate, and bob up and down. Zdog is about 28kb minified. The illustration is made up of dozens of different shapes so there's no WebGL context, no shader compilation, no worker threads. The animation loop is small enough I can write the entirety in:

const frame = () => {
  if (spinning) illo.rotate.y += 0.005;
  t += 0.03;
  orbit.rotate.y -= 0.018;
  illo.translate.y = Math.sin(t) * 3;
  illo.updateRenderGraph();
  if (!reduce && visible) raf = requestAnimationFrame(frame);
};

Two other important things to know: • An IntersectionObserver exits the loop if the canvas is out of view, so people at the halfway point in the case study page don't have to pay to render an animation they can't see. • If prefers-reduced-motion is enabled, it only renders one still frame. The component is created with next/dynamic and ssr: false so the server only sends the browser an empty box of the right dimensions, so the page doesn't jump.

Mosaic: Vanilla-Tilt, hover only

The tiles that make up the project grid all lean over to the cursor a little, using the Vanilla-Tilt library. They're tilted 6 degrees with a scale factor of 1.01. I only import it for devices with a real hover that don't have the preference for reduced motion enabled, and destroy each instance on unmount.

const ok =
  window.matchMedia("(hover: hover)").matches &&
  !window.matchMedia("(prefers-reduced-motion: reduce)").matches;
if (!ok) return;
import("vanilla-tilt").then(({ default: VanillaTilt }) => {
  VanillaTilt.init(nodes, {
    max: 6, speed: 500, scale: 1.01, glare: false, perspective: 1200,
  });
});

On a phone, these tiles are just tiles. A card tilting when you swipe your finger on it is a bug, not a feature.

Terminal demos in CSS

Two of the projects are a command-line tool and an API, so a screenshot of a terminal isn't readable at the width of a phone. So I made a text version of the demos. Each line is a span with its delay, character count and typing time all stored in CSS custom properties, and the typing is done using a clip path animated in steps.

.term.armed.play .ln.ty {
  animation: at-ty var(--t, .8s) steps(var(--n, 20), jump-none) var(--d) both;
}
@keyframes at-ty {
  from { opacity: 1; clip-path: inset(0 100% 0 0); }
  to   { opacity: 1; clip-path: inset(0 0 0 0); }
}

A couple lines of JS connects everything together, and when a third of it is visible, it plays it all back once and then rests on the transcript. Without JavaScript, or if your device has the preference for reduced motion enabled, you just get the transcript. Every line is from the project itself: the CLI demo used its recorded frames, and the API demo used the actual response from the public repo's stub.

Browser windows

The web apps are shown in a browser window drawn in HTML, with a tab per screen, with the background color set on the app. Screenshots in WebP from the running product, not mockups, lighter than video, and always crisp no matter what size they are, no one standing around waiting for the loop to get to the part we want to see.

Structured data

Every page has a Person JSON-LD record for the person's name, role, education, location, profiles (so that it's always the same person), which everything points to via @id. The homepage is a ProfilePage. This is far and away the cheapest of the bunch and they are the ones that are most closely read by search engines and AI assistants.

Things I removed

There used to be ThreeJS, React Three Fiber, the particle web worker, GSAP with ScrollTrigger and the three responsive helpers that allow it to stand up on a phone. Not a single one of these is on the visitor's page. The new site animates three things that each have a reducedMotion path. The old one animated basically everything. I wrote a whole post about how it did it, for 15k particles forming a letter.