About VoidSource
An independent AI evaluation site, written and edited by one person who actually builds with these systems for a living.
Hey, I'm Boris
Senior AI engineer based in Germany, with a PhD in physics (RWTH Aachen / Forschungszentrum Jülich, generative models for medical imaging). I build production AI systems end to end: LLM applications, agentic systems, and data pipelines with real validation. The skill is picking the right approach per problem, not picking sides.
A lot of my client work right now is document intelligence for German enterprises: turning messy real-world documents into structured, validated data. That's one area I happen to do a lot of, not the whole story. The constant is making AI actually behave in production, whatever the use case.
What VoidSource is
I built VoidSource to keep my own AI evaluations honest. Once you evaluate these systems for paying clients, you stop trusting vendor leaderboards and start tracking your own numbers, and a public, comparable version of that didn't exist. So I made it public.
What you'll find: benchmark scores aggregated from primary publishers (LMArena, LiveBench, SWE-bench, Aider, and official model-card numbers where they are reproducible), pricing pulled from provider documentation, and decision guides I write by hand based on what actually changes a model choice in production.
What's automated, what's human
Numerical data (benchmark scores, pricing, context window, modality support) is collected and normalised by an automated pipeline with quality checks; new model releases are reviewed before publication. All editorial commentary and decision guides are written by me. Nothing here is generated and published without a human deciding it should exist. If a number looks wrong, it probably is, and I want to know.
What's not on the site
No paid placements, no "sponsored model" rankings, no affiliate fees. That keeps the editorial side uncomplicated.
Work with me
I take on a small number of consulting and build engagements (see services). If that's relevant, the contact pagegoes directly to me. That's also where corrections, missing models, and methodology questions land. Otherwise the data is free; use it.