0:16 All right, good afternoon. My name is Rebeka Melber. I go by Cam, and I am the director of programs at Istari Digital. I’m here today to talk a little bit about how we can make AI-designed hardware safer. So, a little bit about my background, I’ve spent a majority of my career working in aerospace. It’s a very no-fail industry, for sure. So, if I have some extreme views on some things, you’ll have to forgive me for that.

0:49 Something that’s a little bit different about me than a lot of the predecessors that have been up here today is that I have more of a background in infrastructure. Which is not really sexy, and not something that people get excited about, but I think it’s pretty cool. And so, hopefully with this presentation today, I can get you excited about it as well. Part of having a background in infrastructure, or in aerospace, probably is going to make the next piece not very surprising.

1:14 A few months ago, our CTO came up to me and said, “I want to vibe code an airplane. Let’s do this.” And I laughed, because I thought that’s kind of crazy. Maybe parts of it, maybe pieces of it, certainly the software that supports it, we’ve been doing that for a long time now. But, to truly sit and vibe code an airplane as somebody without a mechanical background, that seems nuts.

1:36 He didn’t laugh. He’s serious. He’s really challenging us to go outside the bounds and figure out, how do we actually make this happen, and how do we do it safely? So, in today’s talk, what I do really want to focus on is we know that AI can build, we know that AI can design, but do we trust what is actually going to come out of it on the other end, especially on something as high fault as an airplane?

2:07 We have been talking about this all morning. The future is here, the future is now. It is a great time to be alive, especially if you’re in aerospace because there’s a lot of planes being designed and built right now. Cloud code, CloudX or CodeX is not just making development faster and making processes better, but it’s opening the aperture on who actually has a shot at doing designing.

2:32 It’s really changing the game, but it’s also shifting skill sets. So, instead of hands-on keyboard being the power like it was over the past decade, today it’s hands-on the problem. If you truly are an expert and have that domain expertise on how things work, you can tell it what needs to be built. And that’s a big change from what we’ve seen over over the past couple of years.

2:59 And it’s really starting to take the power back from developers and put it back into the experts’ hands, into all of our hands and everybody that’s in the room here this week. So, if it’s that great, why isn’t this just the end of the story? AI is actually making a lot of this worse. Don’t throw anything at me yet. I promise it’s going to come back around.

3:24 Why is it actually getting worse? Our infrastructure is broken. Our infrastructure is broken and our data isn’t ready. And the idea of all of that being in place, even figuring out how to be compliant as companies, we have a long way to go on all of this. We’re not really ready to be AI ready, and we all say this all the time, garbage in is garbage out.

3:47 But I like to say garbage in is really just faster, automated, bigger garbage out on the other end if we’re not ready for it. I promise I won’t spend too much time on this because infrastructure is my favorite thing to talk about, but this is really where it starts because our infrastructure is broken. And to me, infrastructure is what enables everything that we’re trying to do. Every idea, whether it’s small or large, needs good infrastructure, needs good data to start from.

4:21 We think across all of our teams. I know this isn’t just an aerospace problem. We have multiple teams that have to figure out how to work together. They’re probably not sitting on the same systems. Even if you’re a smaller startup and you only have 30 people, I bet you have more than one or two AWS environments or Azure environments or maybe a couple servers sitting in a room somewhere.

4:45 Even if that’s not the case, are all of your teams sitting on the same tools? Are the tools that they’re working on Do they work together well? Do they interoperate? What about the other teams and the other organizations that you work with? What if you have to work with different companies? This problem exponentially blows and it gets harder and harder to understand where is our source of truth?

5:08 Where is the good data that our elements can pull from so that we can start to have confidence in what spits out on the other end? The other piece of this is compliance. There’s a lot of scary stats on this about where we’re failing at accountability and how often people are using it and what it really looks like within within the different teams. The stat that scares me the most is this one.

5:40 Over 40% of AI-generated code is failing compliance checks. Now, we’re talking about code. How do we fix this? We can do bug fixes, we can do updates. Maybe worst case scenario, we have to put a PR update out there to say sorry about that, our code wasn’t that great. But it can be fixed. What happens when we go into the hardware? Hallucinating material property can’t get patched.

6:17 A simulation that’s converging on the wrong answer because the input data wasn’t great or it wasn’t from the most recent version doesn’t get a hot fix. When we’re talking about hardware and we’re relying on AI and something goes wrong at some point, it still gets built. And then it can fail. And to me, that is the scariest part. There’s a story in a book Project to Product.

6:46 It’s a little bit older, it’s about 10-year-old book. But it was a story about a major aerospace company who for the first time was releasing an airplane that had parts that was controlled by software. And so there was a lot of fear about that. So when the software teams and the engineers came up to them and they said, “We’re done. Everything is good. We have full trust.

7:08 Our code is excellent.” What did they do? They put all the engineering leads on that plane for the first flight. And given everything that you know and everything you’ve seen and every photo with a sixth finger on it, would you want to get on that plane? I don’t know if I would want to. Now, before I thank everybody for not throwing anything at me cuz I know I just beat up on AI for a while and AI is our friend.

7:39 But it’s really not all AI’s fault. A lot of it actually our fault. AI is doing exactly what it’s supposed to do. It’s moving quickly, it’s generating fast, it’s exploring broadly to produce an output, but it’s also filling in gaps. But what is it filling in the gaps with? That comes down to us. We have to build a trust layer so that it’s filling in gaps with the right information.

8:07 We talked before about the power that’s shifting from developers to domain experts. That expertise is what is needed so desperately as we’re continuing to make this transition because they have to understand what are the guardrails that have to be in place. What is the compliance that must be met? We have to give AI boundaries and a way to know which version is real, which rules apply so that the inputs are trustworthy because then the outputs are also trustworthy.

8:43 The problem isn’t AI and the answer isn’t just don’t use it yet. The answer is that it needs a trust layer. We need it to have a trust layer so that it can watch our back as we’re expanding the envelope for how we’re continuing to use it. Now, let me circle back around. I started this off by saying my CTO challenged me to go five code in airplane.

9:09 How do we do this? How do we get there? Probably the biggest place to start. There’s a lot less crazy things to do than an airplane. Why not sneakers? Why can’t an athlete who knows the performance outcome that they want be able to design what their product looks like? Or a material scientist being able to describe the properties and then AI generates the geometry. The ability to to design what we want is out there.

9:40 We’ve seen it all morning about where we we’ve been able to do this. The real question is, are we giving it the right foundation to give us what we really need? I haven’t talked about Astari yet. I’ve spent a lot of time talking about the problems that we’re facing today and where AI is making things maybe not as much better as we think that they are. But this is really where Astari comes in.

10:10 So, Astari is a product company. We’re a self-hosted software that gets to installed as part of your infrastructure on your systems. It lives locally. It becomes the trust layer as your foundation. And what does that really mean and how do we do this? We have to start with the data cuz the data is the most important thing here. You already have all the data in the world and you have the domain expertise.

10:39 But we do need to make sure that all of the data that you have is AI consumable, that it is auditable, and that it is normal. We talked a little bit about that source of truth. Are you pointing it in the right direction? Again, I’m going to keep coming back to you and you have trust in what’s coming out on the other side. So, let’s get into how we do this a little bit.

11:04 The first thing we do is we extract data. Your data today is trapped. I’m sure many of you are using a load of different types of softwares and those softwares are going to save your data most likely in very large, very proprietary data types. AI doesn’t really know what to do with that. So, what Astari is going to do first is it’s going to extract those large models down into artifacts, Vendor neutral, non-proprietary, open, machine-readable, EULA-compliant artifacts.

11:40 It’s going to allow us to do a couple different things with this. The very first thing that we are going to do once we extract these these models down to smaller pieces is we’re going to assign UUIDs. The UUIDs are 36 characters that cannot be hallucinated, that every time you change your version, every time you move it, every time you add to it, it’s going to update it for you.

12:10 So, when you go to Uncle Claude, as I like to call him, and ask it to run a new simulation for you, and you’re going to have it show your source of truth, it’s going to point back to a UUID. When it points back to that UUID, it’s giving you the confidence it’s pulling from real data. It has not hallucinated hallucinated anything in between. That is probably one of the most important things that we have been focusing on so that you know where that data is coming from.

12:38 You also have the control of knowing where your data is going. With fine-grained access controls, it allows you to protect your data internal between teams, between organizations, and especially as you continue to zoom out, as you’re working with other companies, being able to share that source of truth without having to relinquish full control of your data, and being able to have the comfort in knowing that you have the access controls in place that supports it.

13:16 Now, this comes back full circle. Why do Why do we care? We’ve broken these very large models down into smaller artifacts. We’ve assigned UUIDs to them and access controls to make sure that it’s not just about the right people getting access to the right data. It’s about AI getting the right access to the right data. How many times do we look at the final version, final final, final this time, use this one?

13:47 Now you can control which is your source of truth, so that you don’t have to worry anymore that someone or an agent is working off of the wrong data. So, to bring this kind of back to what we were talking about today, the question was never can AI design hardware. I think we all know that that’s here. It’s been here for a while. We’ve been taking advantage of that.

14:17 The question comes down to how much can we trust it? And then how do we continue to build that trust to do bigger and more impactful things with it. The first thing is we have to free the data. We have to take all of the great data that you have today sitting in your environments and break it down to be usable by AI. We have to have a way to audit that data so that you know not just where it is and where it’s going, but you can tie back anything that AI has created to those sources of truth to continue to build that confidence.

14:57 Cuz at the end of the day, you don’t want to have to claim that it’s safe, that it’s good. You want to be able to prove by pointing back to good data. That’s what we’re building at Astari. We’re building infrastructure. But it’s infrastructure that lives in your systems. It’s infrastructure that provides trust. It allows you to continue moving as fast as we have been able to, especially over these past couple of months.

15:24 But more importantly, it allows you to move safely. What started today with my CTO challenging me to go and vibe code an airplane, it it still seems crazy to me on a lot of aspects. But the more that we start to break down everything that in my mind is scary about how we do this, the more that it becomes reality. So, I would love to vibe code an airplane with anybody in this room.

15:55 And hopefully one day, we’re going to get there. So, go and vibe code an airplane. Thank you. To learn more about the CDFAM Computational Design Symposium, access the archive of previous presentations, interviews with speakers, and information about future events around the world, visit CDFAM.com.