Projects

Things I've built or shipped recently. Mostly AI infrastructure benchmarking at imec, plus whatever escapes into evenings.

Article - How many devs can you fit on a GPU?

A blog series on self-hosting GPU infrastructure for coding agents, written with my team at imec. We ran real (64) SWE-Bench-style coding tasks through Claude Code against self-hosted open models (Qwen, DeepSeek, GLM, Kimi K3) on rented H200s and B200s, and measured what actually happens when you push concurrency up. Throughput numbers that look great in isolation collapse into long task times, prefix caches evict themselves in cycles (I can't tell you about the DeepSeek's insane KV cache rate!), and the cost math that shows self-hosting is not for everyone's wallet unless privacy is the point. Cool fact: GLM-5.2's resolve rate was statistically indistinguishable from Opus 4.8's. But is a big fat model, no joke.

I scoped the experiments, wrote and edited most of the copy, built the cost comparison, and did the distribution push ( [https://news.ycombinator.com/item?id=49098130](Hacker News) crowd had good feedback)

Read it: aistack.imec-int.com/blog/gpu-self-hosting

The aistack website

The public face of my team's benchmarking work. I co-designed it, wrote the copy, and built it with AI pair-engineering (vibe coded, but thankfully my old HTML, CSS and JS knowledge came in handy for once) then spent longer than I'd like to admit de-AI-ifying and cleaning leftover mess out of the HTML. Static site, deployed over SFTP like it's 2009 (I learned to use FileZilla, felt like Windows XP) and I'm fine with that.

Take a look at: aistack.imec-int.com

This blog

Also a project, honestly. I started it to get my own writing voice back after too much time with AI tools and LinkedIn. Just me frolicking my words.