THAHIR KAREEM · CHIEF OF INNOVATION, DESINGULARITY · DUBAI

Teaching machines to reason about design, not recall it.

Self-taught. One-person technical team. I build the models, the data, and the machines they run on.

452 GB
FLEET VRAM · SELF-BUILT
9.79 M
IMAGES CURATED FOR TRAINING
φ
THE ONLY CONSTANT

ABOUT

No degree. No lab. I started with Python scripts and ended up designing training pipelines, evaluation benchmarks, and a multi-node GPU fleet — because the research needed them.

At Desingularity I lead innovation for a small design group. My job is one question: can a machine hold real design judgment? Everything below is an attempt at an answer.

WORK

One product. One thesis.

DES-MUSE-001

MUSE — a design intelligence model

MUSE emits programs, not pixels. It writes a structured description of a design, and a renderer turns that description into geometry. The renderer is also the judge: if the geometry is wrong, the output fails. No hand-waving, no vibes — a hard gate.

The bet underneath it: design reasoning can be separated from encyclopedic knowledge, and trained on its own.

design DSL renderer-as-verifier geometry gate interior + product design

RESEARCH

Threads I keep pulling.

IDTHREADCLAIM
T-01Programs, not pixelsA model that writes structured design programs beats one that paints pixels, because programs can be checked.
T-02Renderer as verifierGeometry is the one judge that cannot be charmed. Render the output, measure the error, gate on it.
T-03SeparabilityDesign reasoning is learnable apart from knowing every product in every catalog.
T-04Compressed expertiseA small, precise vocabulary can carry most of a design expert's judgment.
T-05Signal theory of beautyProportion — φ above all — behaves like a signal the eye is tuned to receive.

WRITING

Notes from the bench.

FIRST ARTICLE IN DRAFT

Ask me about it

CONTACT

Terse messages welcome.