All right. Hi everyone. Thanks for having me here. My name is Sai. Nouri and I’m co-founder and CEO of C Infinity. C Infinity is a startup that was founded 2 years back. And we are building assembly intelligence. And I want to talk to you about what we’re building here. Show you some really cool stuff in in our mission at Cfini is to start thinking towards the future, right?

0:55 Like now the u the for decades in discrete manufacturing we’ve had this fundamental tradeoff you know where on one axis you have variation which is essentially like products can can be constructed in different variants of each other and often it’s called mix and on the other axis you have the volume which is the number of products that are that are made per individual variation and on one end of the spectrum right Like on this curve here you have the high mix low volume manufacturing.

1:30 This is like the aerospace industry, off-road equipment, semiconductor manufacturing, space systems, defense and basically a lot of the a lot of us in this room are looking at industries like that. And on the other end you have low mix high volume manufacturing like think automotive components fasteners and commodity hardware and things like that. And across this entire curve, the pressures never really change. The products have to get to market faster and unit costs have to keep dropping.

2:04 And especially this is happening as complexity is increasing in different products. And on the on the high mix low volume side, you have the advantage of reconfigurability or configurations and variations. And on the low mix, high volume side, you have automation. The combination of these two becomes high mix high volume manufacturing where you can use the complexity and the reconfigurability that you get in high mix manufacturing with the automation capabilities that you can get with low with high volume manufacturing and to be able to converge that into a lightsabout system is really the final end state that we need to get to but we’re not quite there.

2:40 The challenge is that like products they start as digital entities in 3D systems but then they have to become physical and this transformation of the digital to the physical is is hard and automating that translation is what we’re interested in solving and we’ve seen from all the talks in the last couple of days about how CAD and CE have really really interesting emerging AI and computational layers and we are building that layer for manufacturing where we’re essentially connecting design systems, CAD and PLM systems and manufacturing execution systems.

3:14 By coupling these systems, you can start creating learning loops that are going to optimize your product design. Really understanding how downstream manufacturing is is affecting your product design and how product designs can be manufactured creates this feedback mechanism which we’re calling designing the designer. So to get to our goal of high mix high volume manufacturing you know we started looking at like what does this look like in the end right like in the end state people are talking about a lot of people talking about physical AI and what does that even mean right like my definition of physical AI is autonomous hardware operating in the real world right and one of the advantages of physical AI and it’s a massive opportunity in advanced manufacturing and assembly is one of the clearest use cases in robotics and flexible automation So the challenge is that assembled products are designed in CAD like we talked about earlier.

4:09 But the translation of the CAD to the the actual execution steps is it requires solving a problem called process planning where you have to convert the design structure the u what is also called the engineering bill of materials into the manufacturing structure which is also called the manufacturing bill of materials. And the Mbomb, the manufacturing bullet of materials is constrained by the factory. And the Ebomb represents the engineer’s view of the function of the part.

4:37 Translating one to the other requires solving a process planning problem which is mapping the mapping, reordering and sequencing the parts in such a way that they can be built in a factory and also surfacing issues in manufacturability so that the designer can understand very quickly whether there needs to be updates and changes to the design. Today this is done entirely manually and it’s kind of amazing to see.

5:02 So we work with several OEMs large OEMs you know fortune 100 companies and we look at how they do process planning and they show us their process plans. It’s done on Excel it’s done on PowerPoint right and it’s done manually a lot of back of the envelope calculations and it is it is a problem on which manufacturing engineers in these companies spend approximately a third of their time.

5:26 Some companies talk about how this could this cost them close to $40 to $50 million a year, right? It is a foundational problem in connecting design to manufacturing. And to solve this, we noticed that there’s a lot of similarity with how we’ve engineered software systems, right? And when we when we invented programming languages, we invented programming languages and gave people the agency to start thinking in terms of higher level abstractions, writing code in such a way that like you could actually express the intent of what you want to build and then have an a piece of software called a compiler break it down into lower level execution steps that can be built on an architecture on a computer architecture.

6:11 So we see ourselves as building this kind of compiler for physical AI with auto assembler our product where we’re taking the what needs that needs to be built which is the design the engineering bill of materials and converting that into a sequence of execution instructions you know that can be executed in a factory to create the manufacturing bill of materials. It’s a foundational problem and if you don’t solve this problem, if you don’t accelerate this problem, any attempt to accelerate manufacturing is going to be stalled because you’re going to be bottlenecked by human planning speed.

6:42 So we we think this is a foundational problem that we need to solve to design the designer. Okay. And so to do this right like we we looked at you know how people actually do configuration management and product life cycle management right like and if you look at the configuration space which I will loosely define as the set of all product assemblies that satisfy product requirements you can see that OEMs they manage you know millions to billions of product configurations they don’t actually keep millions of SKUs but they make make it to order in highix manufacturing.

7:15 Thing and each configuration has its own revision history. So each configuration is changing in terms of the you know the parts that that make up that product configuration. I show three configurations over here and a change in one configuration can ripple across other configurations right like and they start affecting each other’s u behavior. So planning every single the planning the execution of every single configuration or revision is a huge engineering burden.

7:45 The symptom of it is in what people call work instruction authoring. They have to write down the execution instructions that get built in the shop floor. But the problem the core problem underneath is planning. How can you translate the engineering structure into manufacturing structure? And a configuration is manufacturable only if a factory constraint plan exists. So how do we solve this problem? So with auto assembler you can it is essentially a very simple you know user interface.

8:15 You upload a 3D CAD assembly. You can take in whatever CAD assembly format you have. As soon as you upload the CAD assembly, we solve we have to immediately solve one of a really hard problem actually. Like an engineering bill of materials is really just a bill of materials, a collection of parts. And from a CAD perspective, it’s a collection of parts that’s just sitting in space.

8:38 CAD does not tell you how those parts are connected to each other. Right? So we have to infer the structure the connectivity of the parts and we do that on up at upload time. So we can identify things like how parts are interfering with each other. How parts are connected to each other. What are the parts that have clearances between them you know and what does that graph structure look like you know and this is this happens to have a lot of really useful you know impact in simulation as well and we can talk about assembly simulation and how to use this structure separately.

9:10 But from manufacturing perspective when you have this you can immediately start seeing things like what are the interferences between the parts. Right? Like so you can we give you that list of interferences and real designs have a lot of interferences between them. When we get data from customers, they’re all messy, right? Sometimes interferences are intended in the design. So we can we can clearly identify what are the interferences.

9:33 We can identify things like where are the similarities between the parts and for and we can identify what are the types of hardware that you have inside your geometry. This is all automatically generated right like from the from the input and we can basically start building sort of building initial intelligence and then the next thing we do is we create a station decomposition in an assembly you typically in a product assembly you don’t build it in one shot you break it down into stations you know where each assembly station builds a smaller sub assembly and you kind of build you you build it incrementally and so each station can represents a small sub assembly that that can then be planned locally and then assembled as the as the station workflow progresses.

10:24 What auto assembler then does is it takes the station decomposition and then constructs a sequence plan. To solve this problem, you have to essentially construct a map of all the possible available motions of individual parts. What is blocking each part? Where are the collisions? Where are the clearances? And organize that into a into a set of feasible steps that can be used to execute the product assembly.

10:49 This is a combinatorial problem. It is a huge space to explore. And so this is where like AI planning really starts shining, right? And we also compute design for assembly metrics where for every step in this process plan that’s automatically generated, we can tell you precisely like how good is that step. If it’s a bolt for example that’s being inserted, we can tell you this bolt is an M5 bolt.

11:13 It has an across flax dimension of 9 mm and it needs this type of socket wrench to be able to install and gives you like very fine level of detail along with clearance analysis about what can be what can be built. And so as a result like you start getting into a very detailed description of the process steps automatically you know and now to this this type of work takes hours and hours to do cuz what CAD engineers do is they use exploded views they use screenshots and they try to communicate the design intent using the those those tools.

11:45 You can also write details about individual manufacturing steps that you want to communicate to the shop floor. I mean this is a supplier purchased part and you want to install that first with pneumatic fittings you know for example it’s just a it’s an instruction that you can provide to the to the user we can incorporate times and then use that to optimize the process plan as well.

12:05 So once you have this this setup of of a full process plan, then the user now wants to start interacting with the AI, right? Like the the the artificial intelligence system generates a sequence of steps, but the user has the flexibility of being able to move those steps around, right? Like you can generate new process plans by by varying what you have. Like I’m going to show you a simple example here.

12:31 You take this hex socket head screw and you want to move it up there. And the planner and the AI starts thinking is that even a feasible step to begin with, right? Like and it’s in it turns out in this case that it’s actually going to be an infeasible step and the planner tells you why that’s in feasible, right? So if you go it cannot assemble the crossbar there because if you go through the sequence of steps, it turns out that bolt is interfering with the ability of the crossbar to come in.

12:59 Right? So the u the the spatial planning is happening essentially in real time right and it is it is interacting with the user and the user can then you know understand why a particular step works or does not work. If there are really really complex material properties and motions, the user also has the ability to override what the planner has stated as a feasible as an infeasible step saying it is actually feasible and then the planner works with it.

So, so what you have here is a really powerful planning system, you know, in and off of itself because it is organizing essentially 3D messy 3D geometry. Remember there’s lots of interferences and it’s working around those interferences and understanding the intent of how the parts are supposed to fit together and putting it in in organized system. So once you once you do the u planning then the next step is to author the instructions.

13:59 All right. So okay let’s see we’re on the same slide here. There we go. Yeah. So what you can then do is auto assembler allows you to create a very detailed set of instructions that a human engineer can look at and go and execute in the shop floor. I mean ultimately we also emit motion paths and so on that can be used by robotic systems. You know, thinking of ourselves as a compiler, but but for we also want to meet people where they’re at today, right?

14:27 Like and when when you’re doing assembly in a shop floor, you want to have like detailed tasks, descriptions, and work instructions essentially that can be communicated to a user or to an engineer to be able to go and fabricate the the product, right? Like and so you can have details like you know like what is shown here the instructions the individual steps are all automatically generated because it’s a consequence of solving the process planning problem and you have a bill of materials or for the tools the fixtures and overall description of the individual steps for for every task and that results essentially in a once you do this work it results in a very nice set of work instructions that you can go and show to a user and they can go and execute it on the shop floor.

15:14 So what we’ve gone through right now is process flow where you start from 3D CAD assemblies get a lot of this intelligence you know about how parts are connected to each other what the purpose of each of these parts or parts are how do you break them down into stations how do you plan each station how do you get those plans together and how do you get a work instruction output out of that but that’s just the beginning right like because the real problem is that there are variants there are product configurations so what happens when we have a product variant.

15:45 Now here is a product variant that that that you can see. Now what happens here is that I don’t want to plan this whole thing right like I want to I don’t want to start from scratch and do everything again. So instead what I’m going to do is I’m going to compare this design against the previous design and I get a diff you know a diff that tells me what exactly has changed in this design and it tells me like u these are the parts that have been added.

16:14 There are some parts that have been removed, some parts have been swapped. We use a combination of machine learning embeddings and similarity along with the graph structures that we compute. You can tell detailed geometry changes as well. We can tell where the geometry changes are going to essentially not lead to any process changes because ultimately sometimes you’re just like replacing this a bolt, right? Like and you can use the same process step but with a different geometry.

16:37 So we’re creating this equivalence class of process plans for different types of geometries. And this diff analysis becomes extremely useful to be able to speed up your incremental planning. And so what we do after that is I’m going to go quickly over here. And what we do is we then say that okay we have the updated the new design. We’re just going to adapt a process plan from the previous process plan that we just computed and that’s it, right?

17:07 Like magically what happens is we just immediately get all the decompositions, the station breakdowns, the individual parts by taking into account the differences between the two designs and you have an updated process plan. The steps are really the same for all the parts that have been reused over here. And now you you now you can see how very quickly you can get into the work instructions and the process authoring for the next for the next update.

17:32 So so that becomes so that loop of essentially going from one design all the way through to process planning and then incrementally compiling or incrementally generating plans is how we think of ourselves as a compiler for manufacturing. You know, and so most manufacturing engineering is variance of work that’s already done before. And so what we do is we turn a validated plan into a reusable asset. And this is useful for design engineers because they can check whether the updates they’re making are going to be manufacturable.

18:07 Then the manufacturing engineers can check whether you know they don’t have to reauthor sequences and process plans. And for the organization, we have everything shared and validated in in one system. So, and because we’re all API first integrations with PLM systems, we don’t really, you know, ask customers to upload their CAD data into our systems and it works, you know, in their four walls. It’s very easy to set up.

18:32 So that’s kind of how we solve this pain point of process planning. And and what I will say then is u very quickly to close out the argument is that the compilation going into incremental compilation by reuse creates this operational efficiency that enables the transition to physical AI. You know so ultimately you can now say if everything is going to be robotics assembly then how does changes in the design lead to changes in the robotic process system and ultimately we want to connect upstream right like we this is creating that intelligence layer it’s my last slide solving the pain point of engineering bill of materials to manufacturing materials addresses the problem that people have today you know but it also creates the path to the future with CAD the robotics assembly and then also creating those feedback loops I talked about.

19:22 One of the things I’m really excited about solving with this is creating designs that are going to consolidate parts with the additive manufacturing. One of the biggest use cases we saw for it value proposition was consolidation of parts. But there’s no systematic way of doing this using the technologies that we’re talking about over here. We can generate sequences or and we can generate clusters of parts that are likely to be combined together into single designs and we design them.

19:48 So with that I’ll I’ll say thank you very much. I hope you learned about how we can do how we can accelerate process planning and if you’re interested in talking about auto assembler please reach out to me. Thank you. Well, thank you all for coming and thanks for having me here today. I think it’s it’s extremely