0:02 So, I’m Chris from Celedon Solutions. I’m going to talk about requirements to manufacturable systems at the very abstract high level where you where your trade space is basically what can humanity achieve. So, you know, we’ve heard a lot of great talks in specific domain areas where, you know, the the design space is is more narrow than, you know, completely what whatever you can do. So, we’re going to talk about sort of where that is at the moment, where it’s going, and what fails right now, and what we should do to fix it.

0:58 So, I’m I’m Chris I’m the CTO of Celedon Solutions. I have a PhD in physics, and instrument design, simulation, topology. So, I I think and I’m I’m sure a lot of people here are AI-pilled as well, but we are really on the cusp of a revolution in engineering. And we’re going from this paradigm of human-driven design to AI-driven design. And what’s really important about that is is human-driven time scales to computation-driven time scales.

1:28 So, it’s not like, “Oh, AI and computers are just better.” Right? Why are they better? They’re better because we can do more in less time, and you’ll see why that is here. So, so what is the bottleneck of human-driven design? Okay? Why is why is this AI future going to be better? Why is it going to be faster? So, right now, when you consider the full design loop of you have requirements to architecture to the actual realization of that in the physical world and then verification of all of it, architecture, realization, documentation to explain that to people using that that realization.

2:14 All of that right now is done, you know, by hand or maybe like 5 years ago, you know, is done by hand. When you introduce agents to a subsection, of this loop, it can speed up that that subsection of the loop. But, once you accelerate that section of the loop, there is still a long time in between actually closing this loop because these steps at the abstract level are still handled by people.

2:44 Recent advancements in LLMs change this. They can now work at that abstract level. They can help to close the loop. So, you know, here you’ll see the the agent move into the center of the loop and it’s it’s changing the human from being intrinsic to each of the steps to managing this entire loop. So, this agent is completing each of these steps. It’s asking questions to the user when it needs information and the user can intervene at any particular point, in this loop.

3:18 Here. So, okay, why is this why is this faster? It’s because human time scales have a floor to them. Humans can’t really operate and process at speeds faster than maybe seconds or minutes even for simple evaluations. But, computers and AI agents can. So, in terms of, computational reducibility here, we have our simplest, easiest to evaluate, objects as constraints and then automated checks and cached results, and AI can really speed up all of that, that is done by humans right now.

3:59 So, instead of, you know, thinking about something, writing out something by hand, typing something into a computer, you can do, you know, six orders of magnitude faster computation by giving AI control over that particular part of the process. So, that is what’s really speeding up this design loop. Is that that computation. Caching and reevaluation capability that the computer is really, take, into their own hands. So, how do you get a system that can do this?

4:32 Because normally people like to read documents, and they like visuals. And while that’s useful for humans, that that form in the document is a very, bad compression of information for the AI to work with. So, you have your AI agents extract all of that unstructured data into a structured representation that can be queried instantly for various types of of problems. So, once you extract that information into this representation, once the design change hits, for a certain class of problems, they can be instantly reevaluated.

5:08 There’s no like hemming and hawing about things. They just propagate through the structure. And that can occur on time scales of of microseconds, it can occur on seconds, minutes, hours. Those automated workflows through that system are important. And then once you have that that structure, you can then make views of that structure for humans or for other AIs. So, you know, that’s very important to sort of have those documents, slide decks, tables, respond to the actual underlying, representation.

5:42 And not sort of, use them as the representation, because they’re sort of bad at information compression. So, using this you can really compress weeks into minutes. So, here, you know, the manual process, serial human hand-offs and execution, there’s so so much time lost in switching costs between these tasks. But when the switching cost is extremely low and the evaluation cycle is extremely high, you can really speed up.

6:10 And this is like weeks to minutes, but really there are several orders of magnitude still left to play with for us humans, I guess, to understand. From a physics-based level, there’s like 10 to the 20 times speed up that is like possible, but I think, you know, realistically in the next decade we can expect maybe a thousand to 10,000 X speed up given compute hardware. So, something else that can be done with this time, we still want we still care about things at human time scales, right?

6:46 We don’t care about necessarily the microseconds, we care about the hours and days and years. But what this microseconds allows you to do is it allows the agent to propose new trade spaces that are evaluated quickly. So, this is what a lot of companies are getting very good at now is evaluating these trade spaces. But on the abstract level, you can you can do this really for the entire system and evaluate completely different domains.

7:12 So, I have a mission that I want to accomplish. Is that going to be an aircraft, a helicopter, you know, any sort of system I can design, a ground vehicle? All of those can be traded simultaneously with this speed up. Another thing that these agents can manage is physics-driven design. So, this is a very simple topology optimization, but basically this this field-driven design of okay, you have a problem spec, I’m I’m to let physics drive the actual solution is going to be very important because not only can these AIs operate at the abstract level, they can operate on the actual data structures themselves.

7:52 Okay, so this is all well and good. What what fails today? Why isn’t this accelerating, you know, every single workflow by by a thousand, 10,000 X? I think the better question is what what is failure? Because there there are so many stories of like, “Oh, the AI hallucinates.” Or “Oh, it makes a mistake. It deletes my production database.” You know, why does that occur? So, you know, why can’t we just go from state X to state Y, right?

8:24 Why can’t we just achieve the the goal? Well, it turns out state Y is a little bit fuzzy. Like, you can’t completely represent a future state that you wish to to achieve. It’s a bit fuzzy. And you also don’t know the exact future state right now. So, that’s also a bit fuzzy. So, both your goal is fuzzy and your current state is fuzzy and your path to that state is also fuzzy.

8:49 So, basically, you can’t precisely model any of this because if you could, that would just be the universe. So, but anyway, so here, requirements define your goal space and what what your system is is sort of an evolution from where you are now to your requirements. Your requirements bound that design. And then when you actually start designing, you are projecting your future state to the goal state that you want.

9:23 Can also do this in in reverse, but generally, you’re building this structure where your system exists within this this goal space. And simulation asks, “Okay, does this match physical physical reality?” And, you know, you can get nice check marks if meets all of the requirements, it’s within my bounds, that sort of stuff. The problem is that reality is is hard. So, your your representation of reality is is not reality.

9:54 So, your your simulations, your goals are going to be off from what actually is is possible or feasible. So, there’s sort of three different errors, I would say, here. You can have a state mismatch at the beginning. You can have bad data. If you have bad data, if you have bad information about your current state, no matter how precise your modeling is, no matter how exact you specify your goal, it’s not going to work.

10:20 Your There’s a goal mismatch at the end. Okay, I need to specify this this goal configuration. Oh, I want to literally just add a table to my production database and call it a day. Oh, well, the AI agent goes in and wipes my entire production database and then adds a table. It’s like, well, I didn’t want that. Well, maybe that’s goal mismatch. You know, you didn’t you didn’t tell it not to wipe it, and it added a table.

10:46 So, you know, you are It is meeting that goal. And this is a big challenge in AI because there are assumed things that humans expect of AI that the AI may not know about. And this is like an alignment issue. And if you if you what is what is happening in AI, the evolution is that this alignment is getting better and better and better and better with every LLM.

11:14 So, it’s it’s very correlated with competency, which is generally unexpected in in AI community. But generally, the better these models are, the more competent they are in estimating what you actually want. So, while this problem is getting better, it is something to keep in mind when working with these systems. And what most of the people in this room I think work on is this unmodeled behavior. So, it is important to have accurate simulations of your world.

11:41 If you don’t, it won’t work. So, you know, and these are getting these are getting very good. But something to keep in mind here is even if you have perfect simulation capability, even if all of your behaviors are modeled, you still need to think about okay, you know, what data do I have? Am I using that data correctly? And is my goal specified exactly, you know, to to the best of my ability.

12:09 All right. And so so how how does this how’s this getting done in in normal engineering? Cuz humans do it reasonably effectively, and we have these checks, simulations, human review. So, we think at the meta level about these goals, right? We’re like we actually ask the questions, is the data good? Have I specified my requirements enough? And when you let AI agents just go out without specifying what you want, without you, you know, communicating to it about what you want, it’ll produce something that that doesn’t work.

12:43 So, the idea is to have the same exact gates that an engineer has, but for the AI agents. One of the issues is that these checks, simulations, and human reviews have been built over decades for humans. So, a big challenge in this AI engineering space is to rework all of these checks, simulations, and human steps geared towards AI. Okay, so where are we at now? I’d like to show you sort of where we’re at with DaVinci, sort of our sort of solution to this this challenge.

13:26 What you’re seeing here is a fully connected graph database of all sorts of things, parts, properties, states, behaviors, requirements, everything. So, this data is all represented in a single graph. So, here are like requirements about your system. These are testable parameters. It can also give you an idea of Does this look right from the humans? You can have block diagrams and tree diagrams here to define all the interfaces of all of your systems.

13:58 This can include the 3D CAD as well. You have states and behaviors. How exactly do I model failure modes, activity diagrams, all that stuff. Fully executable and testable. You can model budgets. You can model timelines. You can model risks. So, all of these are ways to to mitigate failure, I suppose. You can also create views. So, this is a document generated, for example, in a slide deck from the actual underlying data structure.

14:30 An important part about this is not only do you have this data structure for simple execution, you have code objects for execution on the graph itself. So, that’s a big part of this this testing loop is the actual verification of the data. So, this video’s going to continue here and I’m going to play this because I think it’s I think it’s really cool. This is DaVinci actually generating that that graph structure.

14:57 So, just like Ryan showed with a simple input prompt, the agent can create and link all of these structures. So, all of the yellow floating objects here are parameters about the data. So, the sort of ontology that we have is flexible enough to represent any sort of domain specific arrangement or or configuration. So, you could say the domain specific language is encoded into the model itself. So, you can see here the structure that gets built, which I think is is really cool.

15:32 All of the linking of the different components, all of the different interfaces between them, you know, power, thermal, etc. There are some sort of deep level requirements here with constraints, and then it starts building all of the behaviors of the parts here. Which is like fireworks going off in this in this 3D structure, which is awesome. And then also, it’ll link all of the states with transitions all together.

15:59 And then here it’s creating timelines, budgets, and deliverables for for various stakeholders. So, that’s kind of the full the full graph structure, which I I think is cool. That’s what’s happening sort of behind the scenes. Okay, and the end product is is something like this. You have a a design. It’s been, you know, the actual visual here has been text to CAD, but it’s also been it’s also designed to be able to import actual geometries, to call tools that can do the topology optimization, link all of those together.

16:36 That’s a that’s a big issue right now with AI agent tooling is it’s optimized for sort of one specific thing, but, you know, parts come in in steps. So, you need to be able to import those into an assembly, and consider the whole system and mounting interfaces. And also, you know, simulating the full mission sort of end-to-end given the architecture and the requirements. You can think of DaVinci as a compiler for the entire engineering process.

17:02 So, going from requirements, architecture, geometry, manufacturing instructions, product realization sort of is the end goal with checks at every level. So, the road ahead, you we think it at Celedon and I think within the next decade, like agentic AI is going to be capable of designing any system and building any system on the time scale of minutes to 10 hours. So, right now, the rapid concept studies proposals are like basically, you know, completely solved, I would say.

17:40 But, taking it that next step into actual manufacturing, real world tooling, production testing for general systems, I think is sort of the next step here. And thank you. And if you want to learn more, there’s a QR code on the screen. Thanks.