Designing for performance. Why AI is on-prem at architects SimpsonHaugh

‘We've been using GenAI for a long time’

In part 2 of Computing's interview with David Moyes, head of digital at Manchester-based architecture firm SimpsonHaugh, he discusses current uses and future plans for AI, and why public cloud is not an option.

To most of us generative AI may be shiny and new, but architects have been using it for years.

"It just wasn't called that," said David Moyes, head of digital at architects SimpsonHaugh. "It was called computational design, and before that it was called something else, but we've been doing it for a long time."

Like all creatives, architects rely on tools and techniques to come up with and test new ideas. Generative design tools out multiple designs given specific inputs and objectives, which can then be whittled down, analysed and improved iteratively.

The capabilities of these tools are evolving rapidly in the property tech space, said Moyes, and they are now routinely used for site-solve type tasks, but they still require validation by a human architect. "It's great getting all this generative AI information back, but it needs validation. There's a there's a real danger, actually, that people are doing generative AI and not really validating the results."

In a fiercely competitive sector, tools that speed up repetitive processes, allowing architects to focus on areas that demand their expertise are no longer just a nice-to-have, particularly as pressures have grown.

"We're really being squeezed for deadlines, and we're also being squeezed on fees as well. That's common across the industry."

Even the simplest designs require more output from the company, Moyes explained.

"For a drywall package, 10 years ago that was maybe two drawings, now it's 60 or 70. And it's not just drawing, it's digital information as well. Tools like Revit have a lot of metadata - its fire rating, its volume, its costs, its construction sequencing. We're doing an awful lot more with the data now."

All this means that SimpsonHaugh is looking for new ways to automate time-consuming practices, and at infrastructure that can support its needs for high performance and confidentiality, at the right price. This largely excludes public cloud.

Very much on-prem

"We're very much on prem," said Moyes. "In part that's because the datasets that we're dealing with are huge. Cloud is great, but for our use case it's expensive because you end up with caching servers, and you get charged for ingress and egress and all the rest of it.

"It's more cost effective for us to do private cloud and host it internally. That way we can keep the compute and the data right next to each other as well, so that the end user experience is as optimal as it can be. We are also looking at a private cloud for AI."

In pursuit of an optimal user experience at the right price, the company has invested £1.21 million in a full digital transformation to implement VDI and AI throughout the business, with plans to introduce privately hosted AI using private data.

See also The sky-high IT expectations of architects, and how one team meets them

In addition to using GenAI to support the production of designs, SimpsonHaugh is rolling out two other types of AI engine. “Word AI” is the term it uses to refer to tools that transcribe and summarise meetings, write emails and produce site reports and schedules; whereas “Image AI” is used to create detailed images from sketches and text, with AI tools creating options in a "SimpsonHaugh style" to aid communications with clients and collaborators.

Image AI is a new area for the company.

"We've done work with public AI type stuff like Midjourney and Bing Images just as a toe in the water. But we've taken the decision that because of the type of projects that we're involved with and the NDA agreements we've got in place ... doing those types of things in public cloud, even if it's a private channel in public cloud, probably isn't the wisest thing," said Moyes.

"So we're looking at hosting Midjourney ourselves internally, and using tools like [Stable Diffusion] ComfyUI and Automatic 1111, hosted internally, which are then trained on our own datasets for that kind of image generation."

With AI adding yet more complexity to the task of implementing, managing and maintaining a reliable infrastructure, Moyes' advice to other IT leaders in the professional services sector is to do your homework and temper expectations of savings.

"Don't rush it. It's very easy to get drawn in by the bells and whistles of new technology. With most professional services businesses, IT isn't their core business. Pick the right partner. Make sure you evaluate the technology well, don't jump in feet first, and don't expect it to be cheaper than doing what you've traditionally done."

Read part 1 of this interview.