84. 4th of July Replay: The Biomanufacturing Cloud with Synonym's Edward Shenderovich * New Commentary *

[00:00:00] Erum: Hey, Carl.

[00:00:01] Karl: Hey, Ram. Welcome to Grow Everything Live at New York Tech Week.

[00:00:05] Erum: Yeah. Awesome. Let's go Knicks.

[00:00:08] Karl: Knicks, yes. So at Grow Everything, we spend a lot of time thinking about what does it take to turn a biological process and scale it up to industrial scale? What does it cost? What kind of equipment do you need?

[00:00:24] Karl: What kind of process do you have? And there's a lot of unknowns in that. And today we're gonna explore How do you solve those problems?

[00:00:32] Erum: before we get started, we have a few questions to ask the audience.

[00:00:34] Karl: Okay. So how many people here have something to do with biotechnology?

[00:00:39] Karl: Oh, good, a few. Great. Yeah. how many are here who have something to do with AI? And how many of you are here because you're looking for a job because AI displaced you? Oh, good. No one. Oh, okay. Oh

[00:00:53] Erum: no, there's some, there's some people, but don't worry because in the world of biotech, there's a lot of work to be done, and the podcast is called Grow Everything, and everything is a lot.

[00:01:03] Erum: So busy, busy.

[00:01:05] Karl: Yeah. So on Grow Everything, we cover anything that has to do with biology as technology. So that's manufacturing, materials, even space. And manufacturing's a big part of what we talk about. AI is also a big part of what we talk about.

[00:01:19] Erum: Yeah, absolutely. And we are in our fourth year of this podcast, close to 200 episodes, and the question we keep coming back to is that the science is moving very fast, but can we build the factories, the supply chains, and the infrastructure to match that speed?

[00:01:36] Erum: And tonight is what we're gonna talk about.

[00:01:38] Karl: So here's the problem. You're a founder or you're an investor, and you might be in deep tech, but you're gonna build something physical. You have to build physical infrastructure, whether you're building with biotechnology, with microbes, whether you're doing advanced manufacturing, you need physical infrastructure.

[00:01:55] Karl: You have to make things physically, and you need to know how much that infrastructure is gonna cost, how quickly you can build it. You need to understand what those parameters are. But the challenge is doing that and having those calculations can take months and cost hundreds of thousands of dollars.

[00:02:12] Karl: You need the numbers before you can get funding, and that's what we're gonna be talking about tonight.

[00:02:16] Erum: Yeah. Yeah. And it's that gap between having a great idea and knowing that it pencils out is what kills a lot of big projects, and we've seen it happen many times. But our guest tonight built something that we will see working live on this screen, and we will have a real use case.

[00:02:36] Erum: So with that, I will introduce our guest, and then we can, join and welcome him on stage. So our guest is Brenton Alexander. He's the chief technology officer of Roebling. Roebling is a company that is using AI to generate engineering designs and capital cost estimates. Brenton has a PhD in mechanical engineering and has spent his career at the intersection of process engineering and computational tools.

[00:03:02] Erum: So with that, let's welcome Brenton Alexander to the stage.

[00:03:11] Erum: Yay.

[00:03:12] Karl: Welcome, Brenton. We're super excited to have you here on the Grow Everything podcast So just for full disclosure sake, we have had your bosses on the podcast before. They're

[00:03:22] Erum: co-founders.

[00:03:23] Karl: What?

[00:03:23] Erum: They're co-founders.

[00:03:24] Karl: Co-founders and bosses. They're good job descriptions. Yeah. Yeah, exactly. So Edward Sandarovich and, Joshua Laffer have both been on the pod.

[00:03:34] Karl: amazing interviews, so no pressure, especially since you're doing a software demo. So tell us about Robling.

[00:03:40] Brenton:

[00:03:40] Karl: what's Roebling and what problem does it solve, and why hasn't this problem been solved before?

[00:03:45] Brenton: Yeah, so Roebling is aimed at that phase between R&D and FID. FID is final investment decision. this is the phase where you're thinking about taking something, could be in the lab, could be something that has come into your doors from a startup if you're, a Fortune 500.

[00:03:59] Brenton: and you're trying to think about, "How do I take this thing that has, potential interest commercially in something that I do, and I actually bring it to scale?" That whole phase right now, is dominated by essentially manual workflows, consultants. You do lots of engineering work, you do lots of lab work, you pull things together, you build business plans, all to get to that point of actually making an investment decision and saying, "Should I move forward with this or not?"

[00:04:23] Brenton: it's expensive, it's time-consuming, and in the world where AI can help start automating some of these workflows that in the past have been done, quite manually, this is a, real chance to accelerate that phase, and ideally drive better decision-making earlier in the process of commercialization so that, R&D focuses on things that actually have potential to reach commercial scale and be profitable at scale, or investments are made in things that are more likely to succeed because you have better information at that earlier stage.

[00:04:52] Brenton: So that's really where we're focused is, helping innovators, helping new technologies, helping those who are trying to get to scale understand what it actually looks like to do so and then succeed.

[00:05:00] Erum: Yeah. Yeah, that really resonates in the world of biotech. I mean, there's a lot of R&D, there's a lot of unknowns, a lot of costs.

[00:05:08] Erum: So I think we should just get into the demo.

[00:05:10] Brenton: Yeah,

[00:05:10] Erum: let's do it. how do we do this? Should I just give you a use case or how should we do this?

[00:05:14] Brenton: Yeah. Let's, Give me something to try, and let's figure out how to get it on the screen and we'll

[00:05:17] Erum: go from there. Okay. I'm really excited about this because we have been deep diving into the world of bio dyes.

[00:05:23] Erum: So these are dyes that are made through the pr- process of fermentation. And dyes are very pervasive across industries. so our bio dye series, we wrapped up part one last month, and we are having another, part two, in July. So what do you think, Carl? Should we do food dye?

[00:05:38] Karl: Yeah. So food dye biomanufacturing facility.

[00:05:42] Karl: And just to kind of give some background on food dyes, we're in a really interesting place with those. people don't want synthetics in their food or in the products they consume, which makes sense. And the Trump administration has said that they're gonna ban, synthetic dyes by the end of this year.

[00:05:58] Karl: I'm not sure that's gonna happen, but there's-- putting a lot of pressure on food manufacturers. There is a lot of innovation happening in this space, but no one yet has scaled up production for this. So the question is, what does it take to set up a facility like this? Yeah.

[00:06:12] Erum: Well, we're about to find out, so let's work your Roebling magic.

[00:06:15] Erum: Yeah. Let's see how it works.

[00:06:15] Brenton: Happy to dive in. So I've loaded our sort of base screen here. This is- All right ... blank canvas. There's not a lot going on. our Roebling AI's in the corner here, and we're gonna, we're gonna start it cooking. So usually when we work with clients and customers, we provide a bit more context than sort of in a demo like this.

[00:06:30] Brenton: But, we'll just prompt it with,let's build a biomanufacturing, manufacturing facility. We'll see if it can deal with my typos. Yeah. can it produce a color and dye using, Yeast ... like a yeast fermentation? Yeah. Sure.

[00:06:45] Erum: Yeast is a- Yeah.

[00:06:48] Brenton: Ooh let's make it- Red Red Red. and maybe let's locate it in the US Midwest.

[00:06:55] Erum: Yeah. Ooh.

[00:06:56] Erum: Yeah. Let me see ...

[00:06:56] Brenton: and make assumptions for me from here.

[00:06:59] Erum: Oh my god, this is

[00:07:02] Brenton: so great. so this is gonna get started. For those who've used, Claude Code or Codex or any of your favorite models, it's sort of doing the same thing. it's starting with a prompt. It's trying to understand what we're after.

[00:07:12] Brenton: It obviously has information around it being a, a package meant for process simulation, chemical engineering, infrastructure development, and it's gonna start working with me to understand what I'm trying to do. And, you can see on the screen it's asking me many questions 'cause I only gave it, like, two sentences to work with here, about what it is that we wanna do.

[00:07:30] Karl: And is it like other AIs where, like, the more input you give it, the better your output is gonna be?

[00:07:36] Brenton: Absolutely. So I'm gonna tell it to just make guesses for me and go, and we'll see how good the output is. but like any system, we believe really heavily in Roebling at this idea of garbage in, garbage out.

[00:07:47] Brenton: Doesn't matter how good your models are, if you put garbage into them, you get garbage out of them. And so absolutely with these tools, the more you put in, the more information you put in, the more data you load into them, the better information comes out of it, which is a critical piece of how we work with customers, is making sure the information that they have, past and historical information, all that is pulled into the system as context to help the system be better at designing what it is that they wanna build.

[00:08:11] Erum: Yeah. And what exactly are the decision points that their platform is making to build something out?

[00:08:18] Brenton: Yeah, so it has to make a bunch of decision points, and so it's gonna probably take a couple minutes for it to cook for us. and we can see right now it's started to make some for us. it's decided what components it should be modeling for this process that we're creating, so it's gonna have water, glucose, ammonia, sort of standard types of chemicals you'd expect in a fermentation process.

[00:08:35] Brenton: it's starting to think about variables that it thinks are driving the process itself, things like, tank sizes and things. and then it's gonna make decisions around what pieces of equipment to use, how to structure those pieces of equipment, what order to put them in, and then financial decisions around, how much does it cost for labor?

[00:08:50] Brenton: How much does it much it cost for CapEx? All those kinds of things. And so it'll go through all those phases with us and then help us understand what it is that we've designed to build.

[00:08:58] Karl: And how does that differ from what, say, if I was the engineer doing it, I would never be that person, but a h- if I hired someone to do it, how is it different?

[00:09:08] Brenton: It fundamentally isn't. It's following the same process. So, for those who know the Roebling story, we started life as Synonym, a biomanufacturing developer, and we have a whole chemical engineering team, a whole bioprocess team, that, that started with us as we were building our own facilities.

[00:09:23] Brenton: And so we followed these processes. We went through the process of saying, "Okay, how do you design a plant? How do you process design? How do you think about techno-economic modeling?" and the critical piece of how Roebling works is we're not trying to do anything special with the AI. What makes a human work better is also what makes the AI work better, and so we follow the same workflows.

[00:09:42] Brenton: this whole idea of front-end loading or FEL or FEED or detailed design versus basic design, like there's a bunch of different words for this in the infrastructure development space, but it's been around for a century because that's the proper way to follow a process to lower risk. we rapidly accelerate that.

[00:09:59] Brenton: We turn it into more of a continuum as opposed to stage gates. Mm-hmm. But the same lessons apply

[00:10:05] Erum: Yeah. So like then the question is,how does this change the work for human engineers and financial planners and consultants? Does it replace them or change what they do?

[00:10:17] Brenton: It changes what they do.

[00:10:18] Erum: Yeah.

[00:10:18] Brenton: Absolutely. So- So you still need human judgment, you still need expertise to actually make sure what you're getting out of the model is helpful and useful, but it replaces a huge amount of that manual work flowing that's been done in the past. the things of, generating quotes by emailing back and forth, the stuff of building out spreadsheets.

[00:10:34] Brenton: a lot of that manual labor can be rapidly automated in a tool like this, and then the expertise can be used for what the expertise is really useful for, which is making the product better.

[00:10:44] Karl: And what about, when it comes to uncertainties? How does the platform handle those?

[00:10:48] Brenton: Yeah, so that's another critical piece of how we think about things at Roebling.

[00:10:51] Brenton: again, garbage in, garbage out, or, false precision is another big problem in these early feasibility stages. And so the way Roebling thinks about every input to this model is either as something that is known or unknown. And if it's unknown, then it's provided a range or a distribution. And so once this gets, far enough into the model build, we'll be able to look at that.

[00:11:11] Brenton: but its idea is not to give you a number. It's not gonna tell you your facility costs X dollars. It's gonna tell you it's somewhere between here and here with some confidence bound, and help you understand where your uncertainty is coming from, so you can then refine and learn more and drive your process forward to narrow those uncertainties as you move.

[00:11:28] Erum: Yeah. That's how I feel about, like, when, like, ChatGPT first came out and everyone's like, "Oh, this is going to replace marketing people, and, it's going to do all these things." And like we, we work in the world of communications, and one of the things it did solve was, blank page syndrome.

[00:11:42] Erum: like when you're sitting and like, "All right, where do I get started?" I feel like that's very similar. It's like, where do you get started? Okay, you gave me something, and now let me play around with it to make it more like reality versus whatever you're thinking, you

[00:11:54] Brenton: know? Right.

[00:11:54] Erum: Yeah.

[00:11:54] Brenton: So in like the few minutes we've been talking here, right, it has now started to create our process flow diagram.

[00:11:59] Brenton: Oh, nice. It's started to pick equipment. It's starting to link them together. This is part of our process simulation package that's under the hood. so it's starting to build these things on the fly for us that normally would take a couple engineers a fair amount of time to do. Mm-hmm. it can do that for us in the background so we can really focus on the results.

[00:12:14] Brenton: Mm-hmm. So I'll kick it off to sort of the next phases of this as we keep the conversation going.

[00:12:18] Erum: Yeah. Yeah. What is the next phase?

[00:12:21] Karl: Yeah, what is the next phase?

[00:12:22] Brenton: well, it looks like it wants to add some DSP. So- Oh. ... so far it's added our fermentation island- Of course ... and some receiving areas, it looks like.

[00:12:29] Erum: Yeah. Just for- the non-biotechs in the room, DSP is downstream processing. So after you ferments or make the actual product, it has to be refined or filtered in some capacity, and sometimes it costs a lot more to do that than the initial process. So it's great that it's asking you, "Do you want to make sure that the DSP is correct?

[00:12:49] Erum: Let's try to get, it correc- right, so we can get the cost down," because this is what this is about too, right?

[00:12:55] Brenton: Right. So it's clearly making decisions for me on what it thinks my DSP should be. And the expert can come in and look at that and, guide the AI and say, "Actually, you need a, a liquid-liquid extraction step in there or a solvent step in there."

[00:13:06] Brenton: and then the AI can implement.

[00:13:08] Karl: So why wouldn't I do this with Claude Code?

[00:13:12] Brenton: Yeah. Or- No, it's a great question. We get it all the time. you can vibe a factory in Claude Code if you want to. The question is, can you trust the result? Yeah.

[00:13:19] Ed Audience: Yeah.

[00:13:19] Brenton: And, you see this in a lot of other fields, right?

[00:13:23] Brenton: you can put all kinds of things into Claude. It'll give you, very great feedback on how amazing your idea is right before it gives you a really confident response of what it knows. And if you don't have any expertise, it's hard to tell sometimes whether that answer is actually meaningful in any real way.

[00:13:38] Brenton: we've taken a bit of a different approach at Roebling. our approach here is really focused around this idea of marrying the AI with something deterministic. Mm-hmm. And so We don't ask our AI to do thermodynamic calculations.

[00:13:50] Brenton: We don't ask it to do physics. We give it the toolset to run those models- Mm-hmm ... and then it can then iterate on it. And so it's a marrying of that deterministic layer, because at the end of the day, when you build a process facility, it is deterministic. Physics is physics. Mm-hmm. It's not stochastic. Mm-hmm.

[00:14:07] Brenton: And then the AI has the toolset to operate and run that, and then that gets married with that knowledge base, some of which comes from us and some of which comes from our clients and customers from what they have, and you pull all the things together, and that's what enables you to get to an answer that's reliable and trustworthy 'cause you can see where it came from.

[00:14:25] Brenton: Mm-hmm. Whereas if you ask Claude, it'll generate something, it'll tell you where it came from, and that might be a fake reference, and it can hallucinate, and it can be really hard to trace that down. You'll spend more time validating the output of Claude than it would take you to do it manually.

[00:14:36] Karl: So, speaking of that, like what you mentioned, Synonym, the kind of previous incarnation or the,what Roebling has evolved from.

[00:14:45] Karl: What is the data that this model's been trained on?

[00:14:48] Brenton: so first of all, we don't, actually don't train the model. that's key for many of our customers. They don't want their data to be trained by a model for fear of, really important IP leaking. so that's critical to how we work.

[00:14:58] Brenton: The data is made available to the models that we have that have chemical engineering knowledge, have bioprocessing knowledge, and then they have contextual awareness of that data. And so it's built on top of the large language models. We're using Claude right now. We could switch it to Gemini.

[00:15:11] Brenton: We could switch it , to OpenAI products if we wanted to. And then we layer on top of that a contextual system so they have access to the information they need to make decisions.

[00:15:21] Karl: Yeah, and you had said, like, if, you could - code a factory using Claude code, but , the output would be questionable.

[00:15:27] Karl: So how do you make sure the outputs th- that Roebling gives us are accurate and defensible?

[00:15:33] Brenton: Yeah, that's a great question, which I will tell it to keep building. It's sometimes annoying that it- Are you

[00:15:39] Erum: asking the question to Robling? That's cheating.

[00:15:43] Brenton: I mean, it's nice that it keeps checking in with me and- Okay

[00:15:45] Brenton: and I would like it to just do this for the demo by itself. But,

[00:15:48] Erum: Yeah, I actually like that too ...

[00:15:50] Brenton: pa-part of that is, is using this deterministic model.

[00:15:53] Erum: Mm-hmm.

[00:15:53] Brenton: So how do you check it? You literally check it with physics and thermodynamics, right? it says, "I can do A, B, and C," and you say, "Great, run that against the model and see what the output is, and make sure you're correct."

[00:16:02] Brenton: So that, that's a key piece of it. the second piece of it is we have this idea within the tool of expert knowledge. one of the reasons AI works really well in fields like software is, A, it's really cheap to check, but B, there's like a huge knowledge base of software that just exists, right? From open source to other areas.

[00:16:21] Brenton: In process development and infrastructure development, most of that knowledge is locked away 'cause it's basically trade secret. If you've figured out how to save $20 million on your $150 million facility build, generally you don't publish that. Mm-hmm. And so for our customers, it's understanding where they have this proprietary information and coding it into the tool in a way that when the AI is running, it's checked against this validated data, so that if you choose something and they have some internal knowledge that says, "Actually, if you do it that way, you're gonna have problems at commissioning," it gets validated against that

[00:16:54] Erum: Yeah.

[00:16:55] Erum: And I'm curious, like, sort of it's still working right now, right? It,

[00:16:58] Brenton: it is,

[00:16:58] Erum: yes. And so walk us through what's happening in the background.I know we can't see it, but, earlier I was like, "Oh, how does it work?" And, hear about, different agents being used and,

[00:17:07] Erum: I see some of my AI expert friends in here, and they're like, "Yaron, what are you asking?" but yeah, why is it taking a while? 'Cause we're so used to instant answers, and a little bit of thinking from Claude, but yeah, tell us a little bit about that.

[00:17:18] Brenton: well, that's part of it. It is thinking.

[00:17:19] Erum: Okay.

[00:17:20] Brenton: and then in-- as you can sort of see things appearing on the screen, we give it a set of tools, and those tools are what allow it to drive the model.

[00:17:28] Brenton: Mm-hmm. And so it's literally saying, "I wanna put a stream here. I wanna add a process block here. I wanna add this piece of equipment here." They're the same tools that a human user would have and do have through this tool, right? I can go in and I can click and drag things around. Mm-hmm. I can go out to the library and drag things onto the screen.

[00:17:44] Brenton: It's fundamentally doing the same thing. It's just given time to think through and then utilize those tools. And, for anyone who's used these things to build software or whatever, it does chew, it does think, and that's part of the power of it, is it's, able to do so.

[00:17:58] Erum: Yeah. Yeah. And so right now we're just seeing this schematic.

[00:18:02] Erum: It's still being built. Quite a messy one at the moment,

[00:18:04] Brenton: yes.

[00:18:04] Erum: Oh, okay. Yeah. It's still being built, underway. I see little water drops, so people that are listening to the episode, but it's gonna still take some time. Yeah.

[00:18:16] Brenton: Yeah, so it looks like right now, I mean, it's putting blocks on the screen.

[00:18:19] Brenton: It's connecting them together. eventually it'll-- when it finishes, it tries to try to organize them so it doesn't look like such a soupy m- spaghetti mess. but we're at a place right now where it Looks like it's adding utilities and things, so, I don't know, maybe at the next turn I'll tell it to, to hurry up.

[00:18:33] Karl: Can you give some examples of the companies that are using the tool?

[00:18:37] Brenton: Yeah. And who

[00:18:37] Erum: should be using it. Yeah.

[00:18:39] Brenton: Yeah, well, everyone should be using it, that's for sure. There we go. There you go. we have clients that range from startups who are trying to understand this thing that they're working on, what it's actually gonna look like at scale.

[00:18:48] Brenton: That's an important question, obviously. we have Fortune 500, Fortune 100 sort of R&D groups who are very interested in the product. They use it from a different perspective, which is either they have many folks coming through the door, and it's helpful for them to understand what they should pay attention to, and then also for their own R&D needs.

[00:19:04] Brenton: and then we have capital project managers, folks who already have assets who are thinking about annual capital budgets and where they should spend on, improvements, electrification, what's it gonna do for their systems. all those types of customers are in the system itself. in terms of, like, who should use it, we really believe in this idea of better information earlier on will help you make better decisions and ultimately save money.

[00:19:26] Brenton: Mm-hmm. And so a lot of folks put off this stage. They put off techno-economic analysis. They put off FEL1, FEL2 engineering, frankly 'cause it's slow and it's expensive. Mm-hmm. But the end result of that is you make decisions with less information.

[00:19:39] Audience Member: Yeah.

[00:19:39] Brenton: And so if you can accelerate that and turn that into something that takes you days to get answers or minutes to get answers instead of months, turn to something that doesn't cost you hundreds of thousands of dollars to get those answers- Mm-hmm

[00:19:50] Brenton: you should absolutely be doing that more often upfront to make sure what you're investing time, human resources, financial resources in are in things that make sense and are going to scale.

[00:19:59] Erum: Yeah. Why does it take so long to get those cost estimates? You mean traditionally or- Traditionally Yeah. Yeah, traditionally.

[00:20:07] Erum: Not now, 'cause it's happening a lot faster before our eyes.

[00:20:12] Brenton: it's a combination of a couple things. there's a balkanization to how things are done right now, so there's usually an engineering group. They're separate from a costing group. They're separate from the finance group. They're separate from the commercial group.

[00:20:22] Brenton: When the engineers are doing their work, they follow basically a pre-built checklist of like here's the things you do to get to what's called a class five or a class four estimate, and if you don't do those things, it doesn't count. Mm-hmm. So there's a lot of like extra work done. There's a lot of wasted cycles in that.

[00:20:36] Brenton: And then when you deliver your work product to the next group and they say, "Well, that doesn't make sense, you should change this," then it has to like iterate back. so a key component of what we're trying to do is like pull all that together so the technical and the financial and the commercial are talking to each other in the same model- Yeah

[00:20:49] Brenton: so that you don't sort of have that. And then we're really trying to link the inputs to the outputs so that you understand where it makes sense to spend time to dive deeper. Where does it really spend time or make sense to spend engineering time to say, "Go get a new equipment quote"? And where is if I spent four weeks getting that equipment quote but it changes my CapEx 1%, that's kinda wasted effort if you really care about an answer that's plus or minus 20%.

[00:21:14] Karl: And I'm just kinda curious, like what has been the response from the clients that have worked with you guys so far?

[00:21:19] Brenton: I mean, thrilled. Wow. we worked with a client that-

[00:21:23] Erum: Technology.

[00:21:23] Brenton: they had just spent close to half a million dollars and spent six months getting some engineering work, and this gives them a similar answer much, much, much faster.

[00:21:32] Brenton: The idea that you get to... I'll pick on Edward, our CEO, 'cause he was new to the infrastructure space when we started Synonym. early in our days at Synonym, we went through this entire engineering study, and at the end of it we got to a plus or minus 50% estimate, and he was "Why did I just spend three months and however much money I just spent to get plus or minus 50%?

[00:21:48] Brenton: I could've just like made up a number." Yeah. And there's truth to that. Yeah. Right? All models are right, some are... Or all models are wrong, some are useful. Like how much certainty do you need in a model to make a good decision? If you can accelerate that and people are seeing that acceleration, that really is meaningful to making business decisions.

[00:22:05] Erum: Yeah. I will say a fun fact, our episode with Edward is the number one episode for our series. I mean, that's close to 180 episodes, so we'll have to link to that in the show notes. It's a good one. Edward's character, we'll have to have him come up in a minute.

[00:22:20] Brenton: Totally.

[00:22:21] Karl: So can you tell us where we are in the process?

[00:22:23] Brenton: Yeah. So it looks like the process model is built, and it's working on the CapEx and OpEx model now. So if I kinda click through, let's see where we are at.

[00:22:32] Karl: And can you give us a sense in terms of like how it comes up with those OpEx and CapEx calculations?

[00:22:37] Brenton: Y- yeah. So it's running from an internal database of information that we've provided to it.

[00:22:42] Brenton: that information comes from,our, expert history of running Sinonim. some of it comes from publicly available information, some of it comes from places that we, license data from, and we use their libraries. so for CapEx, for example, once this screen ends up loading, after it finishes out the CapEx model, what we're really doing in here is looking at, the list of equipment that's, generated by the process model, all the equipment you would need in the plant, and then it's comparing that against a cost database for all kinds of equipment.

[00:23:10] Brenton: That might be tanks, might be pumps, vessels, and it's calculating a cost for every single one of those, as well as its certainty in that cost. Mm-hmm. OpEx is the same thing. The system knows what my inputs and outputs are. It will compare that against databases that we access in the system for, different types of, commodity prices, and it'll use that to try to generate estimates of pricing, thing from utilities.

[00:23:31] Brenton: so it's drawing from a variety of data sources to generate the inputs to the model, and critically, it'll tell you when it has high certainty, like, "Hey, I know exactly what that costs," and where it says, "I had to go all the way back just to ask AI- Mm-hmm. "... and here's the answer AI gave me," let's put a lot of uncertainty on that.

[00:23:48] Brenton: Yeah. And then you can run that through the model and understand, well, maybe it's plus or minus 50% certainty on my utility cost. How much is that driving the uncertainty in my OpEx? Do I really need to drive that number down or not?

[00:23:59] Erum: I remember seeing some of your, demos through a webinar, and one of the things I found very interesting is that, traditionally, the consultants, they'll do-- run scenarios, and they might run, like, 20 manually.

[00:24:12] Erum: But Roebling runs, like, thousands of scenarios. Is that correct?

[00:24:16] Brenton: Yeah, it can.

[00:24:17] Erum: It can?

[00:24:17] Brenton: Yep.

[00:24:17] Erum: Is that what it's doing right now?

[00:24:19] Brenton: it will when it does the uncertainty, yeah. Yeah. It'll run thousands of scenarios across all the inputs from low uncertainty to high uncertainty, understand, what that output looks like.

[00:24:28] Brenton: It's essentially a Monte Carlo type analysis, and because this is running in the cloud, you can run things in parallel. It can do a whole lot of number crunching all on the fly.

[00:24:36] Erum: Oh, what's a Monte Carlo analysis? Sorry, I don't know.

[00:24:39] Brenton: it's a way of doing- Yeah ... uncertainty analysis. Okay. And so you, you have pro- Sounds like

[00:24:42] Erum: gambling.

[00:24:42] Brenton: You have distributions on all your inputs- Oh. ... and then you sort of pick them randomly and run the model and kinda see what your distribution looks like of outputs.

[00:24:50] Erum: Oh, interesting.

[00:24:53] Karl: Yeah, and you keep doing that over and over again.

[00:24:55] Karl: so where are we in the model? Can you give us a sense in terms of what we're looking at?

[00:24:59] Karl: and, when are we gonna get the numbers?

[00:25:02] Brenton: hopefully momentarily here.

[00:25:04] Erum: It's doing a lot of thinking. This is a lot- I'm sorry. Carl, can you give it a minute? Like- ... it took months to get- Want it now ... this information, and you want it in, like, two seconds? Give us a break.

[00:25:13] Karl: But truthfully, we didn't give it that much data. We're like, "Make a fermentation facility." That's true. We didn't say we want it to be a 10,000-liter facility or anything like that.

[00:25:22] Brenton: Well, what I'm gonna do is switch to another screen- Yeah

[00:25:25] Brenton: just so I can show some numbers here so we- Yeah ... can get to the end point. I

[00:25:29] Erum: think the red threw it off. '

[00:25:30] Brenton: Cause it'll, yeah. It'll take another 15 or 20 minutes to run, depending on how much we wanna let it cook. But- Yeah ... for one that we've already completed, this is what it looks like. This is the equipment list.

[00:25:39] Brenton: so that one I was just showing, it's actually populated now. All the pricing is in here. You can see it's telling me where it came from, references, its uncertainty in this pricing itself. it's giving me understanding of my capital costs. So if I look at my CapEx table, all the directs, all the indirects, all the things that go into actually building a facility up to the full cost.

[00:25:59] Brenton: It's estimating this one at 78 million. if I'm gonna look at my operating costs, it'll tell me full annual operating costs and what that looks like on a dollar per kilogram basis for the actual product I care about. Mm-hmm. Mm-hmm. and all this is, again, live with the inputs to the system. So if I wanna change pricing, I can, and see how these things change.

[00:26:16] Brenton: I can look at financials What it's modeling here is not a terribly attractive investment. It has an IRR of 1%. Mm-hmm. but it's giving that information as I look through, and then finally that uncertainty. So looking at what are the parameters that are really driving the uncertainty in the model, that's this tornado chart looked at sort of on an individual basis.

[00:26:37] Brenton: and then an overall distribution of, - what is this CapEx that's coming out of this model? So this is that idea of I can look at this and say, "Hey, my model's telling me, here in the corner, $78 million, but it's actually somewhere between, 38 and 161," which sounds like an enormous range because it is.

[00:26:53] Brenton: Mm-hmm. But it gives you some sense of, like, the probability range and then some ideas on how to make it better. here it's saying my spray dryer is cre- key driver of uncertainty, and so- Yeah ... let's go get a quote for that and drive that down.

[00:27:04] Karl: That's great. And just to be clear, this data is also for a food dye facility and,we came in with this idea for this fermentation-based food dye facility.

[00:27:13] Karl: You were running it and,basically it would be done in about 15 minutes, correct?

[00:27:17] Brenton: Correct.

[00:27:18] Erum: Yeah.

[00:27:18] Brenton: Yep.

[00:27:19] Erum: Yeah. I'm curious if you could do it, like, backwards. Like, could I say, "I only have, like, a million dollars. What can I do? Can I get far?" Should I just use, like, used equipment or, like, is there a way to do it that way?

[00:27:29] Erum: Have you tried it?

[00:27:30] Brenton: Yeah, I mean,

[00:27:31] Erum: it, it's- not to ask you to do that now.

[00:27:32] Brenton: It's the AI, so you can ask it all kinds of questions, right? Yeah. You can ask it to optimize based off pricing. The forcing function. Yeah. And you can say, "I need this facility to be under $100 million. What can I do?" And it can play with equipment parameters and tell you at 100 million you can make something that produces this much, and this is what your economics are gonna look like.

[00:27:47] Brenton: you can say, "I want this in-" Brazil instead of the United States, what does that mean? It'll change the labor assumptions, it'll change pricing assumptions, it'll give you new numbers. you can ask it to run all kinds of scenarios for you. that's when this becomes powerful, is when you have the model built, then you can start exercising it very quickly, to test all kinds of hypotheses.

[00:28:04] Karl: Yeah, and just to be clear, to-- for the audience and for our listeners, it's not about food dye. This is about basically any kind of process-related facility.

[00:28:13] Brenton: Correct.

[00:28:13] Karl: Yeah.

[00:28:14] Erum: And not just about biomanufacturing anymore, but any industrial manufacturing.

[00:28:18] Brenton: Correct. Yeah. The lessons we learned at Synonym and the lessons we've taken from our prior careers, they're important in biomanufacturing, but they're important in any industrial infrastructure.

[00:28:27] Brenton: And so the platform is made for any industrial infrastructure.

[00:28:30] Karl: So give us a sense in terms of like what's next for Robling. I know you guys have an early access program. Talk about that.

[00:28:36] Brenton: Yeah. So right now we have an early access program, that currently we're signing folks up for. So anyone in the audience or listening that's interested should reach out to us to learn more.

[00:28:44] Brenton: so we're working closely with our initial customers here on implementations for them. as the system gets better and better, we will work with more customers, bring more people into that early access program, and then later this year be opening up access to anyone who wants the tool and wants to be able to drive their numbers faster than they can otherwise.

[00:29:00] Karl: Amazing. And what's the longer-term roadmap look like for Robling?

[00:29:05] Brenton: So it comes down to more information and more fidelity. So- looking at this probability distribution function, this might be useful to make an early-stage investment decision of saying, "Hey, it's worth spending another hundred K to drive this forward."

[00:29:19] Brenton: You need those bounds to narrow. You need to provide more information. And so where Robling is going and where we continue to take the tool, and the company is how do you continue to help people in this continuum of development, this continuum of commercialization, to understand their process better and drive the uncertainty down to the point where you're ready to actually hire, your true engineering procurement construction firm that's then gonna take the plans that Robling will generate for you and turn them into construction-ready deliverables so you can actually build what you need to build.

[00:29:50] Erum: Yeah. Wow, this is so incredible. Something that has taken months can just take a few hours to get started, and maybe not as many people, but still people, human in the loop are still there. But I think we're at the part of our, discussion here where we can open up to audience questions if anyone has a question to ask.

[00:30:09] Erum: We will have a mic going around. Okay - Lisette's bringing the mic. Yay. All right. All right, person I don't know.

[00:30:21] Anna: Hi, Amanda Parks, CTO of M- Dip Materials. I have a question. Does the system incorporate any meaningful models around kind of the socioeconomic, paradigm outside of this facility?

[00:30:32] Anna: So meaning, where your raw materials are coming, from, that kind of thing, where your likely customers are gonna be. We know, all that, like, where are you shipping to, where your competitors might be. Like, is there a sort of a socioeconomic placement, that analysis that c- you could...

[00:30:46] Anna: Like, say you were trying to figure out where you wanted to place this factory.

[00:30:50] Brenton: Yeah, that's right. Yeah. So- It's a great question. It's a great question. And one of the things that we have done with Robling, I'll go back to the process modeling screen here. Mm-hmm. one of the key innovations for how we build is each one of these process blocks, if you wanna think of them that way, those themselves, they're basically spreadsheets.

[00:31:07] Brenton: They're not hard-coded. Anyone can come in, edit them, understand the equations that are going on with them. And so you can model anything you want using a process block, and there's built-in thermodynamic equations and physics equations, so you don't have to, like, recreate those. But fundamentally, you can model anything.

[00:31:21] Brenton: And so we've worked with plenty of customers who they're not interested in the dollar per kilogram at the gate, they're interested in dollars per kilogram at their customer's gate. And so then you add, like, quite literally, a logistics block at the end, and that logistics block is there to represent what you need to do to take the product from your gate to the customer gate, and the inputs and parameters that you model there are gonna be based on where you're located and where they're located and logistics costs and things associated with that.

[00:31:45] Brenton: the AI can help build that block. It can help put the equations in, and then it can help find the data to inform that. Now, if we have good data from one of our libraries, it'll tell you, "Hey, we know really well what it's gonna cost to ship this from, Southeast Asia to California," or whatever it is.

[00:32:00] Brenton: Or it'll say like, "Hey, I don't have that data, so I asked AI, and here's a number, so you can actually get something out of it," but treat that with a plus or minus, eighty percent, seventy percent, whatever it thinks it is, so that you can understand, hey, there's still a number here. I can still understand, this tells me it'll be five to twenty dollars.

[00:32:20] Brenton: If your customer only wants to buy for two dollars, that huge range, it s- sort of doesn't matter. If your customer wants to buy for thirty, the huge range tells you're good, and then you can kind of dive in and say, "Actually, let me go get a quote from a actual shipper to understand the logistics cost," and you can, again, drive that uncertainty down

[00:32:39] Audience Member: Hi, Eric Lima, Cooper Union. I hope I can articulate this question well. So before this was created, and you had to estimate the cost of creating this, you said it would be many months and lots and lots of money, but now it's very much cheaper and a lot faster. So isn't that true for all of the data or a lot of the data, like the OpEx costs, that it's assuming from three years ago, but now AI has completely changed all of that calculation.

[00:33:09] Audience Member: So, how does AI itself affect the accuracy of these, projections?

[00:33:14] Brenton: Ooh, my computer went off. So if I understand the question, it's this idea of, like, how much it costs to operate a plant, how many, lab staff and O&M staff you need- And all of that's changing ... today could be different than what it was three years ago.

[00:33:26] Brenton: Great question. Yeah, it's a great question.

[00:33:28] Erum: Great question.

[00:33:28] Brenton: it's gonna be based on the information you give it. And so for a lot of our customers, the information that goes in there is the information they have from their current facilities. So they have their labor rates, they have that information.

[00:33:38] Brenton: And so to your point, it'll tell them what things look like today, but then you can model that in the tool. You can say, "Well, what if I could actually implement some AI tool?" 'Cause someone came to me and said, "Hey, I have an AI facilities management tool, and that's gonna add X dollars to my OpEx. I need to add that as a cost object, but it's gonna lower my labor rate by three X or four X or something."

[00:34:00] Brenton: Like, you can model all those things. That example is kinda an easy one. You can do that in a spreadsheet, but, like, it's the same idea across the whole range. what if the cost of fabrication went down 20%? the system differentiates between the labor fabrication material costs, so it can kind of piece out, "Hey, for that vessel, how much of that went to the steel versus how much went to the welder?"

[00:34:20] Brenton: Maybe the welder is 20% now. What does that look like? You can play with all those parameters and numbers, and when I play with them, not literally go in and change them, or though you could, you just ask the AI like, "Hey, what if material fabrication was 20% cheaper? What would that mean for this?"

[00:34:34] Karl: up here in front, Lizette.

[00:34:39] Audience Member: So I actually work on a product that's pretty similar to this, but in a different domain. And one thing that I think about a lot is review fatigue. So normally, you're creating this process very slowly, and you have a lot of time to think about every assumption you're making in every piece of the model.

[00:34:55] Audience Member: But when it kind of all, like, plops out of the AI, you kind of are like, "Okay, it looks good to me." and I'm wondering, like, how you approach that problem.

[00:35:03] Brenton: Yeah. We talk about this a lot,the way we solve that in Roebling, or attempt to solve that, our team goes through a lot of, like, structured checking exercises.

[00:35:11] Brenton: It's, like, part of the infrastructure development process, the hazops and hazids. Like, there's all these things you do around that, FMEAs. So a lot of that we've incorporated into the tool. The idea is you have, like, these agents that are doing that work, and the rules that they follow, some of them are industry standard rules, some of them are specific to the customer and client.

[00:35:30] Brenton: So you're trying to automate some of those checks so that the user doesn't have to kinda do the same stuff, and you kinda narrow down what they have to look at. but it is a key challenge is when you find stuff, how do you categorize it? 'Cause you wanna flag everything to the user so they can see everything that they should look at, but at the same time, if you give them a list of 122 things and don't categorize them any way, they just like, I can't deal with this," and you kind of assume.

[00:35:53] Brenton: So we have to, again, through the linkage with the model, understand which of these parameters matter more than others in the answer, and try to order and rank to give the user a sense of like, "Hey, this assumption here doesn't really match, and it really drives your uncertainty. You should pay attention to that one.

[00:36:09] Brenton: Whereas this other one here, that also doesn't match, but it also doesn't really matter," and you kinda shove that to the bottom of the list.

[00:36:13] Karl: Any other questions? I think we have time for one more

[00:36:20] Ed Audience: Yeah, and this,my name's Ed Kosecki. th- in this model, you define colorant as very vague. I mean, normally you would probably be more specific, but w- what do you think it, it took as colorant? Is, I mean, it could be a large molecule, it could be a small molecule, it could be something, I guess it has to assume that can be produced in yeast.

[00:36:38] Ed Audience: did it tell you what it was gonna be?

[00:36:40] Brenton: that's a great question. I'd have to go all the way back to the beginning to know what it thought.

[00:36:44] Karl: we weren't that clear in our specification. Maybe we should've been more, focused.

[00:36:48] Brenton: Yeah, I mean, I can say the first questions it asked me, it's like: What is this colorant? And it gave me a number of potential, beta carotene, like indigo, like a bunch of potential molecules. Oh, okay. Yeah. I guess

[00:36:57] Ed Audience: I, we didn't see them.

[00:36:58] Brenton: The next it's asking me is this intracellular or secreted? It's asking what the feedstock is.

[00:37:02] Brenton: It's asking what co-substrates and nutrients it needs. So it's a- it asked me a lot of questions that I just said make guesses for me and just go.

[00:37:08] Ed Audience: So you said it, guess it, and that was part of the guessing then.

[00:37:11] Brenton: Yeah. So I just said make some assumptions for me.

[00:37:13] Karl: Do you wanna show it?

[00:37:14] Brenton: yeah. If you

[00:37:14] Karl: wanna show it, but it's not all up on the

[00:37:16] Brenton: screen.

[00:37:16] Brenton: I see that now. So yeah, I mean, these are some of the questions it was asking at the top around what is the colorant? How is it secreted? it's asking me about the fermentation. Is this batch fed, continuous? What kind of recovery steps do I need? Are there recycle streams? things about productivity capacity, like do I have data on titers and yields?

[00:37:34] Brenton: it's asking me for a lot of information 'cause it doesn't know, right? Yeah. Now what, what-- when I said just make guesses for me, what it actually decided to do, I don't know. I don't know if it shared that level of detail

[00:37:43] Erum: Yeah, that's really incredible. especially for companies that are younger or maybe non-domain experts when it comes to process engineering, it could really help them just think about what they need to be concerned about.

[00:37:54] Erum: And then like, all right, you're probably not ready to use Robolink. You need to go back and re-understand a little bit or answer the questions and then continue using it with your team, and then you can at least have something a little bit more, fleshed out to talk to other co-founders or investors.

[00:38:11] Brenton: Correct. And that's why we anchor so hard on the uncertainty analysis being the output. Because if you made all of these assumptions and then it came back and told you your facility was 68.2 million, that's false precision.

[00:38:23] Erum: Yeah.

[00:38:24] Brenton: It really is somewhere between 30 and 150 based on the information we gave it.

[00:38:28] Audience Member: Yeah.

[00:38:28] Karl: Yeah. All right. Well, thank you so much, Brenton. This has been one of the most interesting conversations we've had on Grow Everything. Really an amazing, amazing tool. for those of you in the audience or our listeners, if you're interested, you should contact Robolink and learn about their early access program.

[00:38:45] Erum: Yes. and I also wanna thank everyone else that's in the audience. You guys have been great. we will keep the conversation going in the reception area. this has been a lot of fun. More people to thank.

[00:38:56] Karl: Yeah, thank you to NYU Tandon and NYU Tandon at the Yard, especially, Winslow and the lighting team and all of the volunteers that have helped.

[00:39:05] Karl: You guys are amazing. Thank you to New York Tech Week. Appreciate that we were able to do this. And then, of course, thank you to our co-producer, Lizette, and our team at Amplify Media. Yes.

[00:39:16] Erum: Thank you so much, everyone.

[00:39:17] Ed Audience: Thank you both.