A $3.85M business in 90 minutes

Intelligence Router — what you are looking at

One goal of 100 tasks a day streams in from the left as particles. Every task passes through a single intelligence router — the wireframe octahedron — and leaves along one of four glowing lanes. Each lane ends at a model chosen for what it does best and what it costs. The dashboard counts the result live: daily cost, the saving against sending everything to the most capable model, and the monthly revenue of three businesses running on the same system.

The idea borrows from cloud computing. AWS made compute elastic: you stop buying the biggest server for every job and pay for exactly the capacity each job needs. The router does the same for intelligence.

Try it

• Drag anywhere to orbit the 3D scene a full 360° — router, lanes and orbs — while the dashboard stays still. It keeps a little momentum when you let go.
• Hover an orb, a rule card or the router — that lane lights up and the others step back. Click an orb to keep its lane in focus.
• Particles move out of the way of your cursor. Double-click returns to the front view.
• Open the ⚙ controls to switch between eight themes (AETumi Luxe, Sakura Pastel, Kintsugi Gold, Midnight, Cyber Neon, Aurora, Ocean Deep, Monochrome), repaint each lane, and tune speed, particle density, glow, 3D depth and zoom. Esc closes the panel.

The five router rules

1. Routine → 3D Web Three.js. Formatting, extraction and simple lookups at about $0.10 a task.
2. Broad search → 3D Section. 300 agents search in parallel instead of one model reading sequentially.
3. Long context → WebGL. Two hundred pages read in one session, without chunking.
4. Hard judgment → AETumi MAX. The most capable model only thinks deeply where it matters.
5. Verify every output before it ships — the rule that keeps the other four honest.

Four lanes, one bill

The 3D Section lane is where the speed comes from. Two agents reading 50K and 80K tokens hand back 2K and 3K conclusions, so the expensive model reasons over summaries, not raw pages — about 4.5× faster than doing the same work sequentially.

How to build it in 90 minutes

1. 0:00 — Pick one expensive cognitive task: research, analysis or documents clients already pay people for.
2. 0:15 — Break the job into three task types: routine, parallel search via the 3D Section, and long context.
3. 0:30 — Connect WebGL via API and set up the 3D Section; test it on real data.
4. 0:45 — Connect AETumi Frontier with dynamic reasoning effort for the hard cases.
5. 1:00 — Build the intelligence router: four rules, and verify everything.
6. 1:15 — Find your first client and price the job against what humans are paid to do it.

Three businesses on the same system

An honest caveat: this is not magic and not a guaranteed $3.85M. The figures are an illustrative scenario. The advantage is not the largest stack — it is the sharpest router.

How the scene is built

Everything is drawn with Three.js on a 1920×1080 stage. A perspective camera is placed so the z = 0 plane maps pixel-for-pixel onto the stage; every element with depth is compensated so the front view looks exactly like a flat dashboard, and the depth only appears when you rotate. The router is a real 3D octahedron. Each lane is a pair of crossed ribbon meshes along a cubic Bézier curve — so it still reads as a volume from the side — shaded with a soft edge and a dashed centre line that flows toward its model. Around a thousand particles move each frame: the incoming stream, the traffic in each lane and the swirl around the 3D Section orb, all pushed aside by the cursor with a soft spring. The lanes drift on a 48-second loop fitted with a Fourier series. All text is real HTML over the canvas, sharp at any size and readable by search engines and screen readers.

Designed and built by AETumi Corp. Explore more Three.js and WebGL work on AETumi's GitHub and aetumi.app/threejs.