Disclosure & About

Token Crunch is run by Vinzenz Feenstra, a Senior Staff AI Engineer for LLM Applications. He has spent twenty years building systems that have to work — Red Hat's RHEL upgrade tooling (Leapp), foreign exchange infrastructure at Barclays, and antivirus update pipelines at AVG Technologies, where he started as the company's sole technical hire after its acquisition of ewido networks.

At Red Hat he was the primary architect and lead developer of Leapp, the in-place upgrade framework that still ships with every supported RHEL release. He holds the most commits to the framework of any contributor — because he invented it.

The reason this publication exists is straightforward: he kept seeing the same pattern on Reddit and YouTube — everyone immediately jumping to the latest frontier model, expensive and overhyped, while the more interesting engineering work was happening with smaller models that nobody bothered to benchmark seriously. Token Crunch is an attempt to fix that: build real experiments, form real opinions, and give people something concrete to point to when the hype gets loud.

Token Crunch uses an automated editorial pipeline to research, draft, and revise articles. The pipeline is built on Temporal workflows, uses Gemini models for drafting and evaluation, and runs experiments (API calls, research queries, code execution) to back the claims in each article before a word of prose is written.

Every article goes through a multi-pass process: opportunity scouting → outline → experiment planning → evidence gathering → draft → three humanisation passes → a three-evaluator quality gate (anti-generic voice check, claim verification, archetype fit). Articles that fail the claim-verification gate are revised or killed. Numbers and model names that do not appear in an experiment result or a cited source are not permitted in the final draft.

Vinzenz reviews article selection and approves publication. The pipeline assists; it does not replace editorial judgment.

We disclose this because the audience for this publication are AI engineers. They deserve to know how the sausage is made — and because the pipeline's rigour is part of the value proposition, not something to hide.

Token Crunch accepts sponsored articles and sponsored sections within organic articles. All sponsored content is clearly labelled at the top of the article and in any listing. Sponsors do not influence the editorial process for non-sponsored content, and they do not get editorial approval over the conclusions of sponsored articles — only factual corrections.

We are particularly interested in partnerships with inference providers, model labs, and developer tooling companies. If your product is genuinely useful to AI engineers, we are happy to write about it honestly — including its limitations.

If you are interested in sponsoring, open a discussion on the pipeline's GitHub repository or reach out directly.

At the time of writing, Token Crunch does not use affiliate links. If that changes, it will be disclosed on the relevant articles.

Vinzenz works professionally in the AI engineering space. Where a potential conflict of interest exists — for example, if an article covers a product used at his employer — that will be noted in the article.

If a claim in an article is factually wrong, we correct it in place with a dated correction note at the top of the article. We do not delete articles or silently edit away errors. If the pipeline produced a hallucinated figure or an inverted experiment result that made it past review, we want to know about it — open an issue on the pipeline repository.