About

I'm Raphael Friedmann. I work in IT infrastructure in Karlsruhe, Germany, and I started in C++ two decades ago. Most of those years went into the unglamorous part of computing: systems that have to be up tomorrow morning, where the interesting failure is the one nobody wrote a runbook for.

That background is why this site exists. When large language models became something you could actually run yourself, the question that interested me wasn't which cloud API to call — it was what one card in one machine can really do, and where the limit sits. So I built the engine to find out.

What I build

imp is a from-scratch inference engine in C++ and CUDA for a single NVIDIA RTX 5090. It decodes 30B-class MoE models at around 300 tokens a second single-stream, runs 37–72% faster than llama.cpp on dense GGUF, and is the only engine using the card's native NVFP4 4-bit math. Every number is dated and commit-anchored in BENCHMARKS.md, with the command to reproduce it. It is MIT licensed, and it was written entirely with Claude Code.

axo is the side experiment: a neuromorphic learner that trains without backpropagation.

What I write

Two tracks, because two very different people keep asking me things.The engineering log (17 posts) is the build log for imp and axo and how inference actually works at the metal.Understanding & using AI (24 posts) is a numbered course from zero: what an LLM is, what a GPU does, and the honest calls on where AI helps and where it bites. No prior knowledge assumed, and nothing sold.

Everything here is written by hand or reviewed line by line before it ships. Where a claim has a number attached, the number is measured on my own hardware and says which hardware, which build, and which day.

Elsewhere

Code lives on GitHub, the professional profile on LinkedIn, and open-source odds and ends at kekz.org. Mail reaches me at kekz@kekz.org — for questions about the engine, the writing, or work. The writing also comes as a feed, and every article is available as clean Markdown for agents.