FIG. A — NOW SHIPPING
Cohort 04 — tolerance stacks for sheet-metal housings.
48
specimens
7
reviewers
3
rubrics
12
labs
Open brief 2026.04.14
ACTIVE PROJECTS
12
FIG. 01 — EUROBRAIN, 2026
Datasets,
dimensioned by
engineers.
We collect expert annotations on real CAD files from mechanical engineers, structure them against authored rubrics, and ship them to frontier labs. One specimen at a time.
DELIVERED TO —
AVERY · LABS
NORTHFIELD
KAPPA / 9
OSTBLOK·AI
PARALLAX
ATELIER · 04
Six frontier labs. Eighteen production datasets. Rubrics on file.
EXPERT ENGINEERS
284
ANNOTATIONS SHIPPED
38,612
CAD FILES PROCESSED
4,108
RUBRICS AUTHORED
92
§ 03 — IN-FLIGHT
Every specimen, every status.
A small, deliberate palette. Blueprint for what's active, cinnabar for what needs to go back, ink for everything else. Color means something.
Datum reference A missing on three of twelve dimensions.
reviewer · r. duarte · 2026.04.13
TODAY · 2026.04.14
6 OPEN
Active
Cohort 04 · housings
In review
Cohort 03 · weldment cols
Complete
Cohort 02 · GD&T frames
Cohort 01b · tol. stacks
Idle
Cohort 00 · pilot set
§ 02 — METHOD
A short, deliberate pipeline. No black boxes between the engineer and the lab.
№ 01
Experts upload
Mechanical engineers contribute their own SolidWorks parts and technical drawings — files they've actually shipped.
№ 02
Authored rubrics
Senior reviewers describe what a perfect answer looks like before any annotation begins. Rubrics ship with the data.
№ 03
Structured delivery
Annotations are bundled with the original part context and shipped to frontier labs in a typed, validated schema.
The next generation of models will be judged on whether they can read a drawing, not whether they can write a paragraph. That is the work.
We started Eurobrain because the data feeding frontier models is mostly written by people who have never made anything. The result is a generation of systems that can describe a fillet but cannot specify one — that pattern-match on terminology without ever having signed a drawing.
Our contribution is small and specific: we work with practicing mechanical engineers, on real parts, against rubrics they help author. We publish the rubric. We pay on accepted work. We name the reviewer. We treat the dataset as an artifact, not a product.
The first dataset that didn't need three rounds of clarification. We took the rubric, ran the model, shipped on Friday.
§ 05 — PRESS
What people have written.
IEEE SPECTRUM
03.18.2026
Why your next benchmark is a SolidWorks file.
On the difficulty of teaching language models to specify a part.
→
STRATECHERY
02.04.2026
The expert-data thesis, applied.
Eurobrain bets that the bottleneck is rubric authorship, not annotation volume.
→
WIRED
01.22.2026
A small Berlin company is teaching machines to read engineering drawings.
Twenty mechanical engineers, one rubric, and the slow case for boring data.
→
§ 06 — JOURNAL
Notes from the drafting room.
№ 04
CASE STUDY
Sheet-metal housings, dimensioned in nine weeks.
How a single rubric and twenty-four engineers shipped 38,612 annotations to a Series-A frontier lab — without a single round of revisions on the rubric itself.
→
№ 03
METHOD
We don't pay for speed. We pay for accuracy.
On the operating principle behind every reviewer note. Why we hold a four-day clock instead of a forty-eight-hour one.
→
№ 02
SPECIFICATION
Reading a rubric like a machinist reads a print.
A walkthrough of the Eurobrain rubric format — datums, frames, tolerances, and the controlled vocabulary every annotator and reviewer signs against.
→
FIG. 06 — CORRESPONDENCE