Project 03Product
Arch Intelligence
A tool that turns an architect’s sketch into a photorealistic render, built solo for the industry I work in.
- Solodesign and build
- Sketchto render
- AECthe industry I sell to
01
At a glance
The whole project in three steps.
- The problem
Clients can’t read architectural drawings, and projects lose momentum in the gap between the plan and the picture.
- What I did
Designed and built, solo, a tool that turns an architect’s sketch into a photorealistic render of their building.
- The result
A working, shipped tool, tested against real drawings from the industry I market to.
Skills it shows
- Product design
- AI product building
- Prompt design
- Problem definition
- AEC market knowledge
Tools
- OpenAI
- Claude AI
02
The challenge
Why the project had to happen.
I work marketing for an engineering and architecture consultancy, which means I spend a lot of time looking at drawings that clients cannot read. A plan or an elevation is precise and, to anyone outside the profession, close to meaningless. The gap between what a practice has designed and what a client can picture is where projects lose momentum.
Read the full context
Arch Intelligence closes that gap: a sketch in, a photorealistic image out. I built it because I understood the problem from the marketing side rather than the technical one, and because a marketer who claims range should be able to ship something.
03
How I did it
4 steps. Open any one for the detail.
Start from the drawing, not the prompt
Input is the architect’s own sketch; only time of day, material and viewpoint are chosen.
How
The users are architects, not prompt engineers. Input is the artefact they already have, a sketch, and the prompt is assembled from the small number of choices that actually change the output: time of day, material, viewpoint.
Constrain the output to be useful
Weighted the pipeline to preserve the sketch’s structure instead of inventing geometry.
How
An unconstrained image generator invents geometry. What a practice needs is THEIR building, lit and materialised, so the pipeline is weighted toward preserving the sketch’s structure rather than reinterpreting it.
Design for the moment of comparison
The interface leads with the render shown next to the sketch.
How
The value is the render NEXT TO the sketch, more than the render alone. The interface leads with that pairing, because that is the thing a practice puts in front of a client.
Ship it, then use it
Built end to end and used on real drawings to find where it breaks.
How
Built end to end and used against real drawings. Everything I know about where it breaks came from that rather than from planning.
04
See it
Arch Intelligence
Demo — scripted, no live model
Real generations cost per call and this page is public, so the outputs are pre-rendered and the pipeline is simulated on timers. The interface, the presets and the comparison are the real ones.
Pick a sketch, press Generate, then drag the handle to compare.
05
What it achieved
A working tool that takes a sketch and returns a photorealistic render, built and shipped solo.
Aimed squarely at the industry I already market to, which is why the problem definition holds up. It came from watching the industry rather than guessing at it.
The demo on this page runs the real interface against pre-rendered outputs. See the note under it.
06
What I took from it
The render matters most beside the sketch, so the interface leads with that pairing. It is what a practice puts in front of a client.
An unconstrained image generator invents geometry. A practice needs its own building, so the pipeline preserves the sketch’s structure.
Everything I know about where it breaks came from using it on real drawings.