Maria Rampa: Hi, I’m Maria Rampa, and welcome to Engineering Reimagined.
In this episode, Dave Mackenzie, Group Director, Digital & AI at Aurecon, is joined by Andriy Mulyar, Founder and CEO of Nomic, to revisit a conversation on artificial intelligence and how much has changed in just two years since their last chat on our podcast.
They discuss how organisations are moving from early curiosity to practical adoption, and where many companies are still encountering challenges.
They also explore the rise of agentic AI, the importance of building the right data and context foundations, and what it takes to integrate these systems into complex infrastructure environments.
I hope you enjoy this fascinating conversation.
+++++
Dave Mackenzie: Welcome to Engineering Reimagined. I'm sitting down with Andriy, CEO of Nomic, who's coming up amongst this glorious Melbourne weather. So welcome to Melbourne.
Andriy Mulyar: Thank you, beats the negative two in New York.
Dave Mackenzie: We last caught up about two years ago when we got together to talk about AI in engineering. What trends are you seeing that's really shifted since our last conversation?
Andriy Mulyar: It's awesome to be back. Last time I came out to Melbourne, you had folks realising that AI systems were coming. AI systems were going to be embedded into the stacks of every single organisation out there. But they didn't know what to do about it. They were talking about things like getting data ready for AI. They were taking about upskilling their workforce to be able to adopt the new technology. I think over the past two years, we've seen a lot of things play out. Both in terms of the adoption, but also capability improvements in the actual AI systems that people can adopt and the kind of work that AI can do has really just, I think, both surprised a lot of people in the space, but also the adoption of current systems have really fallen behind, I would say, the cutting-edge capabilities of them.
Dave Mackenzie: I think one of the things I've seen is we've gone from this mindset of curiosity around AI to actually people understanding the value that this technology can add. Are there examples you can point to of where organisations have shifted from that curiosity mindset to actually seeing value play out for them in their businesses?
Andriy Mulyar: You have organisations that take a risk-first approach to the adoption and you have organisations that kind of take a 'everything but the kitchen sink' style form of adoption to it. Both of those in terms of specific examples have played out differently across not just like regions of the world, but also like industries. A key example here is across like customer support and customer success. And that's an easy place for AI adoption to happen because a lot of the tasks that there are things like information retrieval, get your data ready in the right spot, have AI help your internal folks generate a response. A really good example of a company that kind of maybe over extended themselves a little bit in the space was like Klarna, they went in and they laid off 5,000 staff internally. This was like 2023, 2024, because they realised they wouldn't need a lot of these folks and then immediately they went in and hired back a bunch of them because they realised AI was not there to be able to offload a lot of the work and tasks happening. But on the other hand, there's organisations that have taken a lot more of a, of a gradual approach to the tech, which I think has been a lot more successful. Where they have spent a lot of time thinking about how do we put the right data layer in place to allow for AI to do the right context building? How do we go in and not do a top-down forcing of the technology on folks, but actually get the right leaders internally that drive that bottom-up adoption that helps steer the organisation into adopting it at the right time and at the right place, depending on their maturity level. And then you also have in the technology space, just full-on deep dives. The place where these systems have been most effective, and I think a lot of the focus of the foundation model companies, the OpenAIs and Anthropics in terms of the data they're training on, which has been coding tasks, there have been entire sectors of the economy fundamentally transformed by the tech in just the last three years. And this is not just simple AI answering a question. This is AI doing real GDP equivalent work, GDP moving work for organisations. In software engineering, for example, just over the last four months, internally in my organisation, we've gone from 20% of the code being written by AI to 70, 80% of code being written by AI. And our software engineers, they're skilled verifiers of the output. That kind of thing is coming for all industries. It's just a matter of the humans being able to verifiably measure the performance of those systems. So I think there's a wide range of how adoption has taken hold of the last couple of years. It's been always a function of a company's data readiness, but not just also data readiness. It's where the focus of big foundation model companies have been.
Dave Mackenzie: Two years ago, we were talking a lot about data readiness and getting data AI ready so you can extract value from AI. And I was just wondering if you could talk a little bit about that.
Andriy Mulyar: Yeah, so the end goal of any AI system being adopted into a workload is that you want the AI system to start work where the human starts work, and you want it to end work where the human ends work. So a very concrete example is, you have some tasks that you do or a workflow that you do that gets kicked off by an email, you know. Maybe a submittal request comes in and that's an instance where work starts and then work ends by depositing a report back or a response back, maybe over email. People will only actually really start adopting the technology if the system that you're using for that technology does the whole end-to-end flow for you because then they spend time verifying that flow as opposed to, you know, moving data around and this is sort of what it means to be data ready. In order for an agentic system to do an end-to-end workflow for a human and allow that human to be an orchestrator of it, not in a verifier of it as opposed to somebody that is moving data around and kind of filling in gaps for that AI system, what you need is to have a data foundation where that AI system can pull in that context at any point in time. So for example, at Nomic, we've built models that allow for AI to understand PDFs, drawings, documents that are highly multimodal. We've built models that allow for AI agents to be able to perform really high-quality search over highly difficult multimodal data. All of that is an exercise in what we call context building, making sure that agents can pull the right context from the right sources.
Dave Mackenzie: We're talking about agents all the time now. I mean, AI and agents can't be uncoupled. What do you think the potential of agents are within engineering and technical organisations? Like one of the things I reflect on is we don't have exquisite data sets. We have lots of unstructured data.
Andriy Mulyar: This is where I think the first sort of real value of agentic systems, especially in design and engineering comes in. Where you can take a lot of the tasks that are subtasks, like for instance, you're about to go in and ship out a drawing to a client. Maybe you're at 40%, 50%, right, and you want to get some initial passes to make sure that what you're delivering is up to spec, right? This is the kind of thing that AI is very good at. It's very good at multimodal reasoning. It's very good at being able to do cross-references between a very complex object and thousands of pages to potentially flag issues. What it's not very good at is being able say, for example, are there clashes happening in some three-dimensional system right now that you're getting two-dimensional slices of in the context of a drawing set? So where this all, I think, starts really fitting in is that when you get this kind of symbiosis.
Dave Mackenzie: What do you see as those sort of cultural traits or those traits of individuals who are really maximising the most of this technology as part of their work?
Andriy Mulyar: Yeah. The thing that really differentiates our most effective people now versus the versus the individuals that are still operating in kind of like the pre-agentic world are people that are really comfortable being orchestrators of systems, people who are more architectural thinkers, people who think about what is the right outcome that I'm looking for and not necessarily what are the nitty-gritty steps to get to that outcome that agents can actually go in and accomplish. The people who are performing the best are the ones that are very much systems thinkers that can do more orchestration and then be able to manage all that orchestration. So not just managing one agent, but multiple agents at the same time.
Dave Mackenzie: When I start thinking more broadly in terms of how organisations adopting AI and agents, I think that's one of the traits that'll actually differentiate organisations. Cause one of things I hear often is that there's some future state where we've all got the most powerful AI tools available. I'm interested in how that collaboration can create space for I guess unique outputs.
Andriy Mulyar: Yeah. Let me do a hypothetical here. So imagine we're in a world in five years, right? Imagine like a mean engineering consultancy. Everyone's gone in and adopted tooling that allows for every single designer, engineer on staff to be using agentic AI that performs meaningful engineering and design work to allow for design work to actually happen. Where does value concentrate in this world? Value concentrates in the talented folks you have and their taste internally to be able to deliver that extra edge of quality, right? There's many ways to achieve the same outcome, There's multiple classes out there of solutions that are all things that a client will accept. What differentiates Aurecon from another firm? It's their taste, the experience of their people that they hired, the things that their teams have done to allow them to deliver that work in a higher quality way. Their better understanding of their clients' real problems versus the problems that their clients might be marketing that they have, right? So in a world where the technology is democratised and a drawing review is perfect for everyone to do, it'll be the ones that have the highest taste in how they actually deliver those outputs and how they pick which issues they're actually going to focus on or which issues they're not going to focus on the AI services, for example. I think about in general technology adoption in this manner, which is in a world where everyone has it, it's an equaliser, right? What is that big differentiator? And to me, frankly, it's taste. If I was to define the word taste, which is a very vague term, taste, another way to frame it is quality bar. What is your quality bar for delivering out an asset? What is your quality bar for delivering something with a ribbon on top of it that you're proud to put your name behind it, right? That's the thing that really differentiates somebody who's doing the work just to do the work and clock out at five, or somebody who really deeply believes in the work that they're doing and they're executing on. And if you as an organisation are able to find folks and make folks enjoy working with you that have that level of taste, that have that high quality bar because they're really proud of the underlying work they're doing, that is a thing that differentiates you in the environment where every single one of your competitors has access to agentic systems that can do that work at that level of quality.
Dave Mackenzie: You see it play out in multiple levels now where like the investment we make in our people, the expertise that we have, continues to be the driving force behind what separates great outcome from okay outcome. You touched on quality just then. One of the things I often hear is, these AI systems can be prone to make an occasional mistake or an error, is that at odds with quality? How can these systems increase quality? I'd just love to get your take on what's the feasibility in practise of using AI to fundamentally improve quality output.
Andriy Mulyar: Think about how current project execution happens, right? You bring in a project into the organisation. Requirements are clear, the legal stuff cleared up, execution is happening. Who's actually executing on this project? What's the very last step that happens before you put the ribbon on top and you ship it out to a client, right, that senior principal sits down and verifies the output.
Dave Mackenzie: We think about AI, we think about jobs, we think about people and I think you touched on a few of these in terms of systems thinkers, architectural thinkers, orchestrators, skilled verifiers. What are some of the key and critical skills that our people or people in other technical organisations should be developing for the future?
Andriy Mulyar: Just like any transformation, there's not going to be necessarily a change in the need for very skilled junior people, there is going to definitely be a change in the requirement set for what they need to execute on. And we're in this very strange time, I think, right now, where the capabilities of the system far exceed the preparation of folks that have been training for the roles that they're in, where a lot of folks, it presents, number one, a very difficult time for folks that are not ready to adapt, but also presents a lot opportunity to folks that can very quickly adapt, and it's the people who previously might not have had opportunities to get into certain roles, because they have a skill set that actually is very well-inclined to being things like an architectural thinker and an orchestrator, and they can have better job opportunities.
Dave Mackenzie: So we're thinking about readiness in terms of getting our people ready and the change that's coming. I look at where AI in software engineering was a few years ago, and I've seen massive, massive improvements there. We have AI in design engineering now and for other technical organisations. Where will we be in two years? I'd love to get your thoughts on what the next frontier is in terms of where we're headed.
Andriy Mulyar: The reality of the capabilities of AI systems for the last few years is that the places where you've seen the most acceleration in terms of the capabilities have been in domains where the inputs are text heavy. So as an example, in 2023, you probably tried ChatGPT and the very first example is, have it write an email for you. And you're like, wow, that’s kind of quality for me to just copy paste and make a few edits in, right? That probably saved me five or 10 minutes. That AI is operating over top of a textual domain and then the next place you saw probably a lot of capabilities and maybe a lot of you have tried this, which is AI for what we call like vibe coding, being able to go in and build software products, be able to build features and software products. And while code might seem something that's very different than like writing an email, the actual underlying data domain that it's operating over top of is still text. You're just outputting text, which happens to be code that then compiles into a feature or an application. What's the big difference between that in design and engineering? In design and in engineering, you're working with highly multimodal assets, the inputs to a design and an engineering workflow is not just text, you might have a spec, you might some email that's coming in and asking you a question, maybe an RFI, but it's never just text. There's an attachment, so that attachment is a PDF that might contain a sketch of a detail, right? That attachment might be a product spec that, let's say a contractor is trying to sub in instead of something that is currently being recommended for use in the project by the designers or engineers. And over the past couple of years, this has actually been a sub domain of data that the foundation model company is not focused on developing AI system capabilities on multimodal inputs. And this is actually where we at Nomic are seeing a big acceleration in both capability adoption, but where we're building our own domain-specific models to close that capability gap. So at Nomic, when we say we build agentic systems and accelerators out in construction, what we really mean is, we build models that can understand those highly multimodal inputs and provide that sort of set of lenses for the AI foundation models to be able to perform at a higher quality rate and a lower error rate over top of engineering discipline and subdomains. So a really concrete example here is, say, a submittal review, right? A submittal occurs when a contractor has a qualm or thinks that they have a better idea of what should be going into the project when it's being executed in comparison to the recommendations of a designer or engineer. And you might say, you know, you want this sort of bolt in this sort area of your building. And then the contractor says, these bolts are cheaper, and they can get them more readily available in this region where it's being executed. Let's do this instead. And they'll go in, they'll send the submittal package. And then that's basically a question with a bunch of PDF attachments. The designer and engineer has to figure out, hey, can I actually use this, right? Is this going to break code requirements for the project? Is this going to harm other phases of the project that maybe that contractor's not privy to? And they have to take in all this multimodal context, information that's happening in Teams messages and emails. Information that's coming from the raw drawings, those multimodal inputs, and then make a very critical decision that might lead to tens of thousands of dollars or hundreds of thousands of dollars of down-to-line damages if a decision is wrong. And it turns out that AI systems for the last couple of years, up until I would say frankly four or five months ago, the error rate was way too low on these kinds of inputs because you had this highly multimodal data modality. But things like our parsing models internally at Nomic close that gap. So what you should think about when you think about how are these systems going to change my workflow as a designer or engineer or somebody executing on the delivery of a physical asset is more data modalities that are multimodal and tasks over top of multimodal data modalities are going to start being solved by these systems. And not just reading documents, not just reading multi-modal PDFs like drawings, but actually interacting with multimodal systems, being able to go in and take an action on your behalf in Revit. You find an AI agent will find an issue in your drawing, and you'll be able to say, hey, I want to fix that in Revit, and they'll be actually take that action for you. This is not like any science fiction. We have systems internally on the R&D side of things where we can go in and actually do that, and they can make a call in to Revit with an agent. So these kinds of things are coming. The important part is also always just remembering that even though the AI can go in and produce you something, the engineer still has to spend 30 minutes verifying its output, right, because the AI is not going to put your stamp on it for you.
Dave Mackenzie: Yeah, and I think for organisations like Aurecon where quality is ingrained as part of our culture and way of working, I think we're really set up to leverage that as AI continues to get embedded in our firm and organisations like ours. We really look to AI to be something that amplifies that trade for us.
Maria Rampa: You’ve been listening to Engineering Reimagined, with Aurecon’s Dave Mackenzie and Andriy Mulyar, Founder and CEO of Nomic.
While technology is advancing rapidly, today’s conversation highlights one consistent theme: the organisations that will see the greatest benefit are those that combine strong technical capability with the right culture, human skills, and standards.
If you enjoyed this episode, hit subscribe on Apple or Spotify and don’t forget to follow Aurecon on your favourite social media platform to stay up to date and join the conversation.
Until next time, I’m Maria Rampa. Thanks for listening.
Embedding AI into complex infrastructure environments
Aurecon’s Group Director Digital & AI Dave Mackenzie sits down with Nomic CEO Andriy Mulyar to unpack how AI has evolved to execute practical impact across industries and the skills people will need in the future for success.
As AI becomes more capable, the differentiator isn’t access to tools but how organisations use them. The ability to think systemically, orchestrate outcomes and apply professional judgement is becoming more critical than ever.
The conversation dives into the rise of agentic AI, the growing importance of strong data foundations, and what it really takes to embed AI into complex, real-world environments like engineering and infrastructure.
While technology is advancing rapidly, today’s conversation highlights one consistent theme: the organisations that will see the greatest benefit are those that combine strong technical capability with the right culture, human skills, and standards – and when these elements are embedded into AI agents and workflows to drive both efficiency and innovation.
