AI is getting wild.
Over the weekend I uploaded 3 photos of my house to an AI architectural modeler, typed one sentence, and walked away just to see what I would get since it allowed that method of modelling as well I could have sketched it and modelled. A few minutes later there was a two-story house with a front-facing gable, an attached garage, a porch, board-and-batten siding, gutters, downspouts, a driveway, and a couple of trees in the yard. Then I circled a spot on the wall and typed “there isn’t a window here” and it removed it. There were a couple other funky things, but again, only 3 photos as a lazy test of the tool.
The tool is Plamo from Illoca. The short version: it’s like a smart SketchUp that you talk to.
What Illoca is
Illoca is an AI startup out of San Ramon, California. The founders came from Google DeepMind, the Autodesk AI Lab, and Tesla’s BIM group, and Bessemer Venture Partners backed them. They have two products. Tracing Paper is the 2D side, aimed at sketches, markups, and bubble diagrams. Plamo is the browser-based 3D modeling workspace, and that’s the one I played with.
AEC Magazine called what Plamo does “vibe modelling.” You direct the model at the level of intent instead of drawing every line. The part that matters more to me is the output. It isn’t a mesh. Plamo produces editable boundary representation geometry with parametric logic behind it, so you can change a dimension after the fact without the whole thing falling apart. That’s the difference between a toy and a tool.
A word on “vibe.” Some people use “vibe coding” and “vibe design” as a put-down, as if working from intent instead of specification is cheating. That framing is wrong imho. Vibe design is conceptual design. We have always vibed. A pencil on a napkin, a form massing study, sketchup modelling, a form exploration that gets thrown away by lunch. That was never a shortcut around the process. It was the first step of the process, and every architect and engineer I know still works that way. The only thing that changed is the napkin now hands back editable geometry.
What I asked it to do
I uploaded three exterior photos and gave it this prompt:
review all provided images and produce a highly-accurate and high-fidelity reconstruction of the building.
That’s it. No dimensions, no floor plan, no site survey.
The first thing it did was build itself a plan.9+ steps, tagged by phase: establish foundation geometry, build primary wall volumes, construct roofs and porch, model openings and trim, apply facade articulation, complete structural interiors, add restrained site context, then two review passes. You watch it check them off one at a time while tool calls scroll by in the sidebar. Program Read All. Design Plan Init. API Read Doc. Program Edit. It reads like watching a coding agent work, because that’s essentially what it is.
The whole thing took a few minutes.
What it got right
More than I expected. The model came back with a two-story main house, a front-facing gable roof, a set-back attached garage on the right side, a porch canopy with posts, beams, railings, and steps, recessed glazing with frames and mullions, a sectional garage door, vertical board-and-batten on the body and horizontal lap siding on the front gable, corner boards, trim bands, fascia, gutters, downspouts, ridge caps, flashing, and roof vents. It added a paver driveway, planting beds, and a couple of trees. It wasn’t 100%, but pretty amazing for 3 photos.
It also told me what it guessed. Estimated from the photos: 8.8 m house width, 10.8 m depth, 2.85 m storey height, 6.4 m garage width, 31 degree roof pitch, 2.35 m porch depth. Every one of those shows up as a slider in the Tweaks panel with a min, a default, and a max. Don’t like the roof pitch? Drag it.
And this is the line I keep coming back to. It said interior rooms were intentionally not invented because no floor plan or interior references were provided.
An AI that admits what it doesn’t know and refuses to make it up. I’ve worked with humans who couldn’t manage that.
What it got wrong, and how fixing it works
It put a window where there isn’t one. So I grabbed the markup pen, drew a loop around the spot on the 3D model, typed “There isn’t a window here,” and hit send. Gone.
Then I selected another window and typed “change this to a 14 inch high 50 inch wide non sliding or opening window.” It swapped it out. Each change gets its own version. I was on V11 by the time I was done, and every version is a thumbnail you can jump back to.
The Layers panel is where I started to take it seriously. It’s organized the way a real model should be: base structure (foundation, floor plates, interior structure), building envelope (exterior walls, roofing, roof trim, porch structure), facade details (glazing, window and door frames, exterior doors, battens, horizontal siding, architectural trim, gutters), and site context. Not one big blob called “house.”
There’s an Export button. I didn’t get that far.
Where it stalled
I ran out of credits.
The free plan gives you 1,000 starter credits. The Explore tier bumps that to 3,000 a month. Team plans start at $59.99 per team per month with multiplayer and SSO, and there’s an Enterprise tier with custom pricing. My handful of prompts and a dozen versions burned through the free allotment in about an hour. Fine for a test drive. You’ll need to pay to do real work.
My take
I’ve written before that the highest value, lowest risk place for AI in our world is the QA and audit layer, not open-ended geometry generation. This doesn’t change that. If Plamo generated a retaining wall from a photo, I’d still want an engineer to check every dimension before it went anywhere near a plan set. But for massing, early concept work, and getting an existing building into 3D fast enough to start a conversation, this is already useful. And it’s a beta.
Plamo is an AI-first tool, not a legacy modeler with a chat box or mcp server just bolted on via a plugin or api. That’s the part I expect to spread. We’ll see a lot more of these, for a lot more design tasks, across a lot more disciplines. Architecture first because that’s where Illoca started, but mechanical design, structural, site, and the rest aren’t far behind. The economics of building one of these just got a lot cheaper than building another SketchUp.
This is also a different thing from most of what gets called AI in our industry right now. Most of that is an MCP server or a drop-in agent wired onto an existing product. Useful stuff. But that’s instrumenting software that was designed 20 years ago for a human with a mouse. Plamo was built from the start around an agent that plans, calls tools, and writes geometry. The chat isn’t a feature. It’s the product. That’s the thing to pay attention to.
Illoca is also honest about where it fits. The pitch is that it feeds Revit and other CAD/BIM tools rather than replacing them. That’s the right pitch. The last thing any CAD manager needs is another island.
If you want to try it, it’s at illoca.com. Bring photos of something you know well. That’s the only way to judge how much it’s guessing.
Cheers,
Shaan
Sources:
illoca.com
Bessemer Venture Partners on Illoca
AEC Magazine on Plamo
