Skip to content

Work

21 projects

The Thoughts

  1. explaining thingsJUL 2026How do you explain a complex, honestly boring technical subject so someone leans in instead of glazing over?
  2. adjacency is not accessJUN 2026A floor plan tells you which rooms touch. It will not tell you which rooms you can actually reach, and the gap between those two drawings is where the architecture is hiding.
  3. buildings that respondJUN 2026Right now a building is a fixed guess. Someone chose the ceiling height and the window size years before you arrived, and you live inside their averaged assumptions whether they fit you or not.
  4. experiences are dataMAY 2026You learned what a room does to you by being in rooms. Nobody sat you down with a dataset.
  5. comfort as dataMAY 2026You don't walk into a room and average your experience. You walk in and the thing that's wrong is the thing you notice: the glare, the echo, the cold draft on the back of your neck.
  6. when the tool scores peopleAPR 2026I build things that give a room a score. The obvious next question, and the one I would ask me, is what happens the day somebody uses that score to decide who gets the room.
  7. latent spaceAPR 2026Machine learning borrowed the word space from us, and I would like to check what it did with it.
  8. drawing as interfaceMAR 2026For most of my life the drawing was the output. You designed the thing, then you drew it to prove it existed.
  9. behavior information modelingJAN 2026For thirty years we've been getting very good at Building Information Modeling: every beam, every duct, every clash, tracked to the millimeter.
  10. heritage meets new techDEC 2025The usual choice with an old building is a trap: treat it as a museum piece no one may touch, or drop a glass-and-steel box in the middle and call it modern.
  11. neuroaestheticsDEC 2025Neuroaesthetics is the unglamorous name for a question everyone already feels: why does one room settle you and another keep you on edge?
  12. evolutionary searchNOV 2025Evolutionary search is what I reach for when a problem is too tangled to solve head on: describe what good looks like, then let a population of designs breed, mutate, and compete until something clever falls out.
  13. physics solversNOV 2025A physics solver is the most honest collaborator I have: it doesn't care what I intended, only what I actually built.
  14. computationOCT 2025Computation is not the same as using a computer, and separating the two took me longer than I would like to admit.
  15. what an llm actually isSEP 2025A large language model is a machine that guesses the next word. That is not me being kind to you, that is the whole thing.
  16. generative aiDEC 2023Generative AI made the inspired gesture free. Anyone can produce a thousand renders before lunch, and a render is a promise, not a product.
  17. extended realitySEP 2022I started in a lab strapping headsets onto students to see whether a lesson lands harder when you can walk around inside it.
  18. connecting the dotsJUN 2020When the art class was asked who here dislikes drawing, mine was the only hand that went up. Awkward.

lEgoarCh

How can AI turn a text prompt into a LEGO set that actually snaps together? 

TYPEComputation & ResearchAWARDJURY AWARDTEAMMACAD GENERATIVE AI · WITH CHARLES ABI CHAHINESTACKFLUX.2 KLEIN · TRELLIS-2 · LORA · LDRAWLINKSGITHUB (opens in new tab)BLOG (opens in new tab)
1 / 11

WHAT

A render is a promise, not a product: you cannot snap a JPEG together on your living-room floor. lEgoarCh takes a text prompt and returns a LEGO Architecture set that is digitally verified buildable: AI imagines it, deterministic code makes it snap together, brick by brick, out of real catalog parts. Built with Charles Abi Chahine, end to end as a pair; I stitched the stages into one flow, prompt in, set out, and gave it its interface and its story.

WHY

The inspired gesture is cheap now; anyone can generate a thousand renders before lunch. The interesting problem moved to verification: what does it take to turn a generated image into something that provably holds together?

HOW

  1. A LoRA fine-tuned on a 40-image dataset teaches FLUX the LEGO Architecture look; the prompt becomes a styled render.
  2. TRELLIS lifts the render into a 3D mesh, and the mesh is voxelized into brick space.
  3. An optimizer places real catalog bricks into the voxels, enforcing connectivity, support, and perceptual color accuracy.
  4. The set exports as LDraw, the format the brick world already speaks.

WHAT CAME OF IT

The most instructive moment was a failure: an intermediate model came back connected and supported, yet not legible as architecture. Structurally sound and visually wrong is still wrong, so legibility joined the constraints. The final pipeline was benchmarked on three buildings.