MIJULABS
AN AI DATA LAB FOR TASTE

Human taste,
at training scale.

Miju Labs turns the judgment of working designers, illustrators, photographers and writers into preference data, critiques and evals for generative models.

Start a pilot How it works
INSIDE A JUDGMENT

Every label comes with its reasons.

A thumbs-up tells a model what someone liked. A Miju judgment tells it why, in the words of someone who does this for a living.

The preference
Pairwise, ranked or scored
The rationale
Written, specific, reusable
The rater
Discipline, seniority, calibration
PAIR 12 / 40Brief · Book cover, landscape

Which output has better taste?

Output A: a busy lamp-lit workshop scene
AOutput A
Output B: a quiet dusk horizon with one lit window
BOutput B
Preferred

Rater’s reasoning

B holds one idea, the lit window, and gives the title somewhere to sit. A is lovely but has nowhere for type to go.
Rater 0417Book cover designer
WHAT WE MAKE

Three ways to put real judgment into a model.

Every label is made by someone whose eye has been tested on real work.

Two framed paintings on a gallery wall, one lit by a shaft of sunlight
01 · PREFERENCE DATA

Pairwise & ranked preferences

Side-by-side and best-of-n judgments on your model’s outputs, from raters matched to the medium.

A sketchbook with pencil notes on a desk by an open window
02 · EXPERT CRITIQUE

Critiques that say why

Written rationales on composition, type, color and voice, plus rubrics your team can reuse.

A long sunlit gallery corridor lined with framed artworks
03 · EVALUATION

Held-out taste evals

Expert-scored benchmarks that tell you if a checkpoint got better, or only got different.

THE NETWORK

Rated by people whose taste is their job.

Every expert is admitted on shipped work, calibrated against their peers, and paid like the professional they are.

Apply to the network
  • IllustratorsIMAGE
  • Type designersLETTERFORMS
  • PhotographersIMAGE
  • Art directorsLAYOUT
  • Motion designersVIDEO
  • WritersCOPY
HOW IT WORKS

From brief to dataset.

  1. i.

    Brief

    You tell us the model, where it falls short, and what “good” means to your users.

  2. ii.

    Calibrate

    We draft a rubric with your team and a panel of domain leads, then test it on gold examples.

  3. iii.

    Judge

    Matched experts compare, rank and critique. Every label carries a written rationale.

  4. iv.

    Deliver

    Data, agreement stats and rater notes land in your bucket, with a lead on call.

FIELD NOTES

Writing from the lab on taste and the data behind it.

A flock of birds over a valley at dusk, one breaking away
METHOD · COMING SOON

Agreement is not the same thing as taste

A lamp-lit artist’s workshop at blue hour
THE NETWORK · COMING SOON

Why we pay raters like consultants

Field notes, once a month.

One email when there’s something worth reading. No roundups, no filler.

START A PILOT

Give your model
a better eye.

Start a pilot Join as an expert
GET IN TOUCH

Start with one pilot.

Labs: tell us what your model gets wrong about good work. Creatives: show us the work you’re proudest of. Either way, a person reads it.

I am

Tell us about the model and where its taste falls short. We’ll come back with a scoped pilot.

We reply to every message.