Kog
The landing page of a startup that runs AI models up to 30 times faster. A very technical subject, made clear in seconds through illustration.
Visit the site- Client
- Kog, with Figue
- Role
- UI designer and illustrator
- Year
- 2025
- Duration
- 1 month
- Scope
- Landing page and illustrations

Overview
Kog is a French startup that speeds up language model inference: its engine generates up to 3,000 tokens per second, where ChatGPT produces around a hundred. The team of PhDs and GPU engineers earned the French Tech 2030 label. They needed a landing page that lived up to their technology.
My role
I worked on this project as a freelancer for the design agency Figue, over one month. I designed the page and all the illustrations. I didn't build it. The visuals below come from my Figma design.
Challenge
Kog works in a highly complex field: communication between GPUs, parallelism, latency measured in microseconds. Every concept was hard to depict. The page had to stay credible for an audience of engineers, while getting the key point across at first glance: with Kog, AI answers much faster.

Solution
Each section relies on an illustration that turns a technical idea into a simple image: two chips facing off, a work loop, a memory diagram. The dark purple world, with its streaks of light, gives the product a cutting-edge feel without falling into AI clichés.

Section by section
- 1
Showing the loop
A coding agent works in a loop: it drafts, lints, tests and refines. The page puts the two rhythms side by side: one 5-minute cycle on a standard stack, versus 50 cycles in the same time with Kog.

- 2
Drawing the engine
Each strength of the engine has its own visual: bars for speed, a connector for drop-in integration, a memory diagram for hardware optimisation, and the partnership with AMD.

- 3
Proving it with numbers
The proof fits in four cards: 3.5 times faster than the industry-standard tool, validated by AMD's engineering teams and featured on their technical blog.

- 4
A simple ending
The page ends on a direct question and a single action: request access to test Kog on your own agent.
