Skip to content

Blog

Knowledge & practice

Field reports on AI automation, enterprise integration and sovereign data processing, written from our own measurements and published with the numbers attached.

Technology

Measure first, recommend second: our test bench on Hugging Face

Before we recommend an open model to a client, we measure it on one machine with an RTX 4090. Code, cases and raw rows are on GitHub, and since 10 October 2026 the result tables and an invoice dataset are on Hugging Face as well. Some things stay off Hugging Face on purpose.

Test benchHugging FaceLocal AI
Michael Boiman19 min read

Technology

Jev against local models: decisions on our own hardware

How far can decision models like Jev be used in a sovereign way, with the data on our own hardware? We compared four open language models and one classifier on our own graphics card with Jev by TypeSafe, with the same questions and all values side by side. Qwen3.8-27B invents evidence least often, Winnow-12B is not measurably behind Jev on any question, and the classifier GLiNER2.5 Decide does no better than the most frequent answer each time. So far models set the reference, no person has judged yet.

JevTypeSafeSystem One
Michael Boiman19 min read

Technology

Laya against Jev: going local is a rebuild, not a swap

Laya is an open model for typed judgments that runs locally and compares itself directly with Jev. We measured both on the same 1,189 cases, on our own graphics card and through the API. Laya is fast and gets within reach on yes/no questions, but not with many options or when it has to pick a value from text.

LayaJevTypeSafe
Michael Boiman13 min read

Technology

Jev in practice: the description beats the model

TypeSafe released Jev on 15 September, a model that returns typed judgments with a probability instead of text. We measured it as a skill router against Claude Haiku, Claude Opus and a keyword search: 1,732 examples, eleven tests, every table in the article.

JevTypeSafeSystem One
Michael Boiman23 min read

Technology

BKS open-bridge: the repo we run BKS-Lab from, now open source

A coding agent starts every session with no memory. BKS open-bridge is a git repository of markdown and YAML that the agent reads first: clients, repos, tasks and what happened yesterday. We run BKS-Lab on it and published it under the MIT license.

BKS open-bridgeOpen SourceAI Agents
Michael Boiman7 min read

More articles in German

Most of our articles are published in German only. The German blog carries the full archive.

Go to the German blog