Episode Content

This article has been prepared for a recent episode recorded for the Digital Construction Podcast, featuring Guido Rivai, Computational Design Lead at Mirvac.


When I asked Guido Rivai for the best advice he has ever received, he admitted he rarely follows advice. His rule is simple: listen to all of it, then make your own judgement, because you are the one who lives with the decision.


It sounded like a throwaway answer. By the end of our talk, it had become the clearest description I’ve heard of how to work with AI.


Guido trained as an architect, became a computational designer, and moved from a design studio to a developer and builder. Each step taught him something about where tools create value and where people still matter.


Tools can now generate almost anything you ask for, so the scarce skill is knowing what people actually need and judging whether the output delivers it.


Hooked by a Better Workflow

Guido never set out to leave architecture. He picked up Grasshopper to improve his own design workflows, got hooked, and kept getting better. In a large office, that meant specialising.

What hooked him was the ability to connect a whole design to flexible inputs, so that when a concept shifted, the model shifted with it.

That led him to a definition I find useful. Computational design is programming for design and construction. It manages the relationships between elements so a workflow can be driven by logic instead of by hand. If you have hundreds of panels, each slightly different, you write an algorithm rather than modelling them one by one.

Realize that this is a definition built around problems. And the most important problems turned out to be at the end of the process, where design meets the people who build it.


Where the Drawings Stop Working

Many architects see computational design as a motor for funky, unbuildable forms. Guido sees the opposite. It often makes difficult things buildable.

He was candid about Revit. It is an amazing platform and the ultimate documentation tool. But at fabrication it often falls short. The person cutting formwork for precast may need unrolled templates, profiles, or even G-code to run the machine. A well-documented sheet is the wrong output for them.

This is a lesson about knowing your user. The fabricator is a different customer from the client reviewing drawings. With the right input, they stop wasting time or simplifying a design that looks too complex.

Seeing that clearly is much easier from inside a business that owns the whole process, which is exactly where Guido went next.


From Design Excellence to Yield

At a design studio, the focus is design excellence and the bulk of the work is iteration against deadlines. At Mirvac, which owns the land, develops and builds, the focus is yield and efficient design. Guido works on the efficient, repeatable side.

That shift changed where he looks for value. In his experience, the biggest opportunities sit at the edges of the design workflow, where it starts and where it ends. So he argues for bringing a computational designer along the whole project, stepping in wherever efficiency appears.

It also changes how you think about software. A company can end up with 30 tools for 30 tasks. Guido prefers a few base platforms extended with custom scripts that could help with specific tasks like solar radiation, pedestrian flows, wind, or compliance checks.

That approach used to have one big weakness: the people who needed the results couldn’t use the scripts. AI is changing that.


The Usability Problem Gets Solved

Guido now uses fewer standard components and vibe codes more of his own. Sometimes he skips scripts altogether and builds a web app.

A node-based script looks alien to most people, and asking colleagues to install software they don’t need is a barrier. A simple web page anyone can open removes both problems.

His example is an urban planning tool for masterplanned communities. Planners trace roads that split the site into lots, apply subdivision logic, and every change regenerates with instant feedback on yield.

So the feasibility risk dropped sharply. Which raises the obvious question: if AI can build the tool, what is left for the computational designer?


Why the Prompt Is Not Enough

Guido has asked himself whether he will still have a job in a few years. He believes he will, and his reason is the most instructive part of the story.

Building that masterplan tool showed him the limit. A single prompt, without careful iteration, produces a tool that answers the literal request and is completely unusable. Getting it right means understanding how planners expect to work, what feels intuitive to them, and refining again and again. AI takes shortcuts to satisfy the request. It misses the subtle parts of a real workflow.

Customers describe their problems well. Turning that into something they will actually use still takes judgement. As Guido put it, it still needs an expert pilot.

The same holds for the rest of AI in architecture. Image generation is fine for early exploration, but once the design is detailed and the visual must match the model, he sees a real risk of wasting time fixing small errors. And architects, like lawyers, sit inside a system of councils, planners and relationships that AI cannot simply replace.


Which brings us back to Guido’s advice. Listen to everything, then make your own judgement. That is how he works with AI: let it generate, then decide what is right for the people who will use it. His message to the next generation is to use AI as much as they can, test agentic systems and automate their workflows, because the alternative is becoming irrelevant. The judgement stays with us.