Building Rhea's Factory: How AI-Designed Enzymes Could Finally Solve Plastic Recycling

Welcome to Just Now Possible with Teresa Torres. I'm Arzu, co-founder and CEO of Riyaz Factory. I'm a molecular biologist and biochemist by training. Had long years of academic training. Eventually find my passion in synthetic biology and circularity and founded Riyaz Factory two years ago together with Mert. Mert. I am Mert, CTO and co-founder of Riyaz Factory. I have been in the technology world for 20 plus years. I'm originally from Turkey, came to the United States Bay Area because I'm a nerd. Worked at big companies and before Riyaz Factory I was at Google for 10 years. Early on project manager and then became a product manager there. And I wanted to be able to use my skills for something impactful. That's when I met Arzu and her vision about changing, disrupting plastic recycling with AI was like really meaningful to me. So that's how we joined our forces. Yeah, I'm really excited to dig into what you're doing because it sounds very meaningful. Arzu, it's clear you have a biology background and a biochemistry background. Mert, do you have a biology or biochemistry background? Are you primarily bringing the engineering mindset? No biology background at all. Biology was probably my worst class in my high school years. I think I learned a lot the last two years thanks to my good teacher Arzu here. But I'm definitely coming purely from computational perspective. Excellent. I like this. This is a good, very different domain expertise but clearly found a way to work well together. Indeed. Okay, so for Riyaz Factory, it sounds like Arzu, you had the seed of inspiration. So tell me a little bit about what does Riyaz Factory do and where did this seed of inspiration come from? We develop technologies for recycling materials and specifically starting with plastics. And we use biology to do that. And our inspiration is coming from the fact that I believe and we both believe we have all the materials out in the world right now. We don't need to dig to generate, we don't need to dig oil to generate new materials. We make new materials and we waste them and we throw them away. And that linearity of our manufacturing and the whole world actually bothered us a lot. This is where everything started. It's why don't we use the materials we generate once very valuable and eventually ending up in landfills? Why don't they become valuable? And that is actually the core of circularity. And that is where our company is started. Let's use the waste and generate value from them. And our tool is biology to be able to do that. Yeah, I like this. Okay, I know from your application, you framed this as today we already have technology for recycling plastic. If I understood correctly, we can only do that so many times. Like each time we use a recycled plastic, it degrades more and more. Maybe we can recycle it two or three times. Is that accurate? That's correct. So the definitions are actually important. So currently the way we call recycling is mostly reusing the materials. Meaning we collect the materials, we melt them down, and reform them into new materials. That causes the quality to degrade by time. Therefore, you cannot really do this so many times. However, in our case and what we do is we are really turning the materials back to their original building blocks. So that it is like to the beginning of their lifetime and they can start being a new material again with the right process applied to them. Yeah, so I want to get a little bit into the science. Enough to give context for when we're getting into what your agents are actually doing. So I would love to hear like with traditional recycling, like what's happening at the like maybe high level molecular level to understand like where that waste is coming from and then maybe talk a little bit about the approach you're pursuing and why that doesn't have the same waste. We have a plastic problems. Everyone accepts this. Yeah. And we can only recycle 10% of the plastic we manufacture and it's been like a hundred years we are trying to solve this problem. Like why? Yeah. The reason, one of the reason is that petrochemical companies are manufacturing plastic. They understand mechanical technologies. They understand chemical technologies. And they were trying to solve this plastic recycling problem with the tools that they understand, which is mechanical and chemical. And unfortunately, those technologies have an inherent limitation to make plastic recycling 100%. That is like the why I was pulled into her vision, which is we need a new tool here, which is biology, because we have been trying to solve this problem with the same tools we created the problem. Yeah. The biggest point about biology in this space, as Mert said, we are getting in, we are bringing biology to the chemical industry. And what biology brings actually is selectivity and specificity. Life can happen in a very small environment within the cells and there are millions, tens of thousands of different reactions happening in a very small space. Chemistry can only do one reaction at a time and cannot generate selectivity because the processes used are either the heat or like high pressures, very general. But our catalyst enzymes, the biological catalysts, are evolved for years in nature to be very selective to different materials and different processes. And that brings a huge potential into the chemical industry, meaning the real problem in recycling currently is the waste being mixed, mixed with different materials. Okay. And that is really difficult to tackle with traditional methods like mechanical or chemical, but biology can tackle it with different catalysts designed for different materials. Yeah, okay. Help me understand, Arzu, you described initially like what the chemical companies are doing, is they're literally melting the plastic, whether that's a chemical process or literally a heat process. Like I'm curious about what is getting lost in that process and then talk to me about how using targeted enzymes, what gets preserved with your solution. With the mechanical methods and applying heat, you actually do not turn them back to their molecular building blocks. Okay. Just simply melt the form and then reform it to the new shape. What it does is it degrades the quality of the polymers, the building blocks of the material, which doesn't let you to generate the quality that is being needed in the industry. What we do is, maybe I give this example that will help better explain. Materials are made out of polymers. You can think of them as necklace of pearls. And what we do is we break this necklace into the single pearl pieces and then re-bring them together. But the traditional methods simply put that together, melt and remold it. And it doesn't really achieve the quality that is needed. Yeah, help me. The raw materials is an output from our process, recycling process. Our plug and play to the existing supply chain as if they are coming from wherever oil, et cetera, like as high quality as they can be. In that melting process, are you losing polymers? Are the structure of the polymers changing? Yeah. With the traditional methods, what is critical is the length of these necklaces. And the traditional methods just break down this chain length and make them shorter. But they don't break them down to the building blocks. They just degrade it. And what we do is we really bring it down to the original building blocks by eating down the polymers and cutting precisely at the right locations and generate the building blocks. Enzyme is part of the solution, but it's not the only part. Okay. So when we talk about, when we realize that, okay, we designed an enzyme, but then we need to have the end-to-end process. Okay. And the enzyme is a depolymerization process. All the activities happening in that process, the material we add, the base that she selects to add comes into play. That's where we are starting to use AI, for example, to optimize not just the enzyme design for the sake of enzyme design, but also how that enzyme will operate in a more user-economic feasible way in the full process. Okay. So if I understand right so far, when a traditional company is applying heat to the plastic, we're shortening polymers, there's some waste. Help me understand why enzymes change that. What is it? I know enzymes are more targeted. Help me understand, like, what's happening. We're not getting that polymer shortening. Like, what is it about enzymes that allow this to happen? What happens is, so the original process, where it starts, maybe this is what we talk, so that it makes more sense what enzymes do in the life. It's more like going back in time of the life of a polymer. So the life of a polymer starts from what is called the monomer, the building block. And then there's a process to generate these polymers and the polymerization. And then you make different materials from these polymers. There are additional additives added during this process as well, which is not that important right now. But what happens is traditional methods take the final product and turn back them to the polymers again, not more than that. But these polymers are not the polymers that are ideal sized, but they are smaller now because of the heat. And what enzymes do is they really turn the timeline back, all the way back to the monomers, meaning that they are now original building blocks. The time zero for the life of a polymer, where they are just monomer, and you can generate as long polymers as you like with your process. That is why the new materials being made from the enzymatic process can be brand new quality. It is like equal quality as the chemicals you make from oil eventually. The first time plastic is being made, first monomers are made. They get combined into becoming polymers. When we recycle, traditionally, we are not getting all the way back to monomers. We are getting to shortened polymers. And what you are doing with enzymes is you are trying to deconstruct the polymers back to their base monomers. For a plastics company, you could input either one and they are going to look the same. Exactly. And that is the value that we bring. Enzymes are basically a byproduct from a bacteria. And they are living organisms. We are not using living organisms in our process. We are not using bacteria. So it makes the scale a little easier. However, they are biological organisms. Therefore, they live and operate in very mild conditions, compared to 1000 Celsius degree heat that we process. We are operating at 60 Celsius, for example. Which means that our energy consumption is much less. Our greenhouse gas emissions are much, much lower. Because when you apply heat to this process, then you can also generate hazardous materials or hazardous gas. Our process is the most sustainable in terms of plastic recycling in the existing technologies. Okay. So I am going to ask a really naive question. So humor me. Are you familiar with how enzymes work in living organisms? Like at a very rough level, they catalyze things. Normally, in plastics, there are no enzymes, correct? Plastics are not living things. There is no... Okay. So really what it is, you are using enzymes in a process to deconstruct plastic into its basic units. Correct. Okay. We take the waste, and I will show you this. We just have some blueberries here. I am pretty sure you have... We take this, when it is becoming the waste, put it in our reactors, together with our enzymes, to break this material back to real building blocks. And I can actually show that building blocks as well. And then Mer, you said something that intuitively makes a lot of sense to me. If we are melting something at a thousand degrees, you are going to get a lot of byproduct waste. There is going to be destruction that you do not want. And I am assuming what is happening with your enzymes is they are deconstructing in a much more targeted way. And that is what is allowing these building blocks to stay intact. Correct. And as you mentioned that they are selective. So if I give enzymes with this plastic that I want to recycle, maybe I have a t-shirt from plastic polyester, but it is also cotton, it has something else. Our enzymes will find the polyester PET, the targeted plastic type, for example, and will only catalyze or break down into monomers. That one, the rest is left over. But when you apply heat, it does not separate. Yeah, it just burns it all up. Exactly. Yeah. Okay. Help me understand. And again, maybe some of this is secret sauce. So that is okay. We are at a high level. But you said you are a reactor. You are adding enzymes to a plastic. I guess in a living system, I can see how an enzyme acts in a cell. What is your process that you are creating this biological process? I am imagining a giant petri dish. No, this is actually a reactor. Okay. Reactor means a closed container where we are able to mix the content so that the reaction can take place much more efficiently. And we are able to apply heat to this container. And that is all. And the heat we are applying is, as Man said, 65 centigrade degrees compared to hundreds to thousands centigrade degrees in the other contexts. Okay. And more importantly, the pressure is also atmospheric pressure, meaning we don't apply any additional pressure. Interesting. Okay. It is happening like in your normal kitchen where something is being mixed and processed. Okay. When I saw this process, I was fascinated because it's so simple. We just put water, plastic, enzymes, and that's it. And he did 265. Yeah. Arzu, I'm curious about where the seed for this process came from. I know there's a lot of, like, creative solutions in science come from, like, somebody from another domain looking at a problem. And it sounds like this might be what happened here. You have a biology background. You got curious about recycling. Traditional recycling has taken a very chemical, physical process. Tell me a little bit about just that seeded discovery for you. I will go back. I will go to a place beyond myself. So this idea and concept is nothing new to science. Science is always looking for new inventions in a much more useful manner. And this has been going on. But specifically for plastic, there have been no effective enzymes found so far until scientists from Japan actually made a discovery from the remainings of a recycling facility. It was a bottling recycling facility, and they were collecting samples of living bacteria within that remains of recycling facility. And they discovered a bacteria that has the potential to break down PET, the material that I showed you here. And that was groundbreaking because it was the first time that a living cell was able to live on plastic. But of course, this wasn't something that could immediately be applied in an industry. It just opened up a space in science. Hey, it is possible. Life has evolved in a direction. If we put the life in a context where there is a lot of this material, then life tries to go and find a way to make use of that material. Eventually, the enzymes responsible for that process have been identified, and scientists start working on that. And this is where I discovered it. I became aware, hey, there are these enzymes. It's so cool. Everyone finds, the scientists find this really cool. But then when I go back to real industry, I didn't see anything happening in that area. And that was the frustration. It's like, why no one is taking this? It's such a cool idea, right? Like, why is this not happening? And that is the beginning of it. Like, hey, I think it really touches something important, but it hasn't yet been applied. And initially, I started understanding why is the case. So my story actually starts by quitting my previous job, which was again for a climate technology startup. But I wanted to dedicate myself to this problem and literally spend maybe six months exploring what is happening in the recycling space right now. Why no one wanted to implement this technology. And I didn't even know what it takes a new technology to be implemented in the market at that time. So the story started that way. The background science done by other scientists, but I myself had been working on different enzymes and engineering them for industrial applications. So that is what I did in my previous jobs and focused on doing that. And that is, hey, there is another enzyme out there that can really help another industry to change the path, but no one was doing it. But then when I start exploring, I start seeing, oh, there are actually a couple of others start doing that, but they are not yet at the market. And seeing at the same time, what was also happening is AI was changing the enzyme discovery field. And it was changing it in a very dramatic way. This was eventually getting a Nobel Prize in a very different context. I was like, I think there's no reason not to get into it now. Hey, we can make these enzymes that nature has evolved, make them even better, apply them in industry and go for it. And I thought, maybe naively, like no one else than me will go and do it because they don't know what the enzymes will do that, what they will do. And that's how the story started. And that's how we joined forces with Matt as well because we both wanted to do something impactful. That is the core of what we are doing, the real foundation. And I wanted to apply a new technology in engineering biology and applying it to a chemical industry. And Matt was completely on board with the idea. And that is how it started. Yeah. Okay. So it sounds like you already had a background in like enzyme design applied to industry, just not this industry. And then it sounds like a second, I'm sensitive to the fact that I'm going to say catalyst, was the Nobel Prize. And if I remember right, that was like in protein folding. Is that right? Correct. Yeah. Okay. So not exactly the same thing. Do you want to talk a little bit about the differences there? Yeah. It's a terminology thing. Enzymes are proteins as well. So proteins are one type of building blocks in life. And what is special about them as well, they are actually linear chains, but they function as folded structures. And that is the key here. And I had been doing my PhD also on this problem, how proteins are being folded to a certain structure and not to the other one and how this process happens. And the structure, if I remember right, impacts its function. Correct. And the thing is, the same chain will never fold to a different structure. It will always fold to the same structure, the same exact chain. But the scientists were not fully understanding how this is really happening, all the biochemical, biophysical properties. Even though we were understanding, we were not able to predict the structure just by looking at the sequence. That was requiring huge computational power that needs extreme amount of power and time. But with the AI technologies that the Nobel Prize came to the AlphaFold, where the linear structure can be given and what it will look like in the folded state could be predicted. And help me understand why this problem matters. Is it easy for us to identify the linear state, but it's not easier? Like, why is this a hard, why is this a problem worth solving? The implications of proteins is in their functions. And the functions, as you said, is correlated from their folded structure. And we are able to get the sequence and the sequence is coming from DNA. I think everyone is familiar with the DNA. DNA has a sequence. And there has been breakthroughs 20 years ago that we were able to sequence the DNA. So we are so easily able to learn what the sequence is, but we cannot easily learn what is the folded structure. And without that, we don't know the function. Most of the time, yes. And normally, until these discoveries, it was taking a PhD student's lifetime to discover the structure of the protein. Literally. And if you discover two, then you are like amazing. And it requires real experimental methods. There are different methods to do so. That is a huge amount of work. And to see how it looks like in another sequence, then you have to start over again. And that is important because as I said, the implications are on the function. If the sequence changes, structure changes, function changes. Yeah, got it. Okay, and then enzymes are a subset of proteins, correct? Correct. Enzymes are biological catalysts. So the proteins that have the function of catalyzing a reaction are called enzymes. Yeah, got it. Okay, so is, I know like in the Nobel Prize, that group is using AI to predict structures from the linear sequence. Correct. And that's helping us understand function. Correct. Are there clear rules? Like you talked about it used to take a PhD for their lifetime to find one structure. Are there, like, is this just a puzzle and like we know the rules, are the rules chemical and we know what structures are possible and it's just like a brute force problem of finding the possibilities given the sequence? Tell me a little bit about why this is so hard. Indeed, it was exactly like this. So we understood the biophysical forces and what it can do, but the number of combinations and probabilities were so high to compute. It was a computation problem. And that is being solved by the neural networks and everything behind that AI tools that I admit that piece how it is happening at the software world. I don't have it with understanding, but what I know is the scientific rules that they base their work on. So the reason why I asked are there clear rules is because I can see very clearly how if you have a large neural net or in this case, a large language model, and it has clear rules of success and you can give it a sequence. We can use this same sort of predictive technology to just churn through. And if it has clear rules to test against, it knows exactly when it's succeeded. Exactly. Okay, so let's get into what you're doing with enzymes and recycling. So it sounds, I'm gonna guess the goal is, we wanna find enzymes that have certain properties that allow us to break these polymers down into monomers with no waste. It sounds like there's two other variables. You're trying to keep temperature low because that's what reduces waste. And you mentioned pressure. So I'm imagining there's a constraint around just the environment to reduce waste. So tell me a little bit about how AI is playing a role in this process. AI is our tool to initially design new and better enzymes. Okay. Initially, we use AI for the design of our catalyst. Yeah. And the vision we have is to take this to even to the more systems level where enzymes are a player within a process. So understanding this different process components with the help of AI is where we want to take it. But right now we are using it to design our enzymes. And maybe to give you a sense of why it matters and what does that mean is, I told you I did a lot of enzyme engineering in my past and I did it in old school. It is now old school. The traditional methods, which is mainly a trial and error. You just test things in the lab. You come up with an idea. You design it and you have to physically build this enzyme and go and test everything, which is time taking and also a lot of resource requiring process. Now we tackle most of this problem computationally using the machine learning tools and design new enzymes that the nature hasn't seen before. Nature didn't evolve to that direction. But machine learning with the algorithms can design these new enzymes and can present them to us and we can test only a subset, a small number in our lab, which accelerates our timelines and also decreases the resources that we use. Yeah. So this is why I wanted to understand a little bit more of the science because I feel like it gives a target for what the agent's trying to do. If I remember right, are enzymes amino acid chains? Correct. Okay. So amino acids trying to guess. So it sounds like you have a protein language model that's been trained on what we know about proteins today. So it has some concept of sequences and what those sequences do. That helps it generate a new sequence. Then there's this second model that you're using that takes that sequence and says, this is what we think the shape is. And then there's a third step, which is given this shape, this is what we think the functions might be. And then there's a last step of, if we did this in the lab, what do we think would happen? Yes, slightly more complicated than that. In the first step, generation can be also multiple steps, meaning, okay, am I in the directionally right direction in terms of a protein being a stable protein using maybe protein language models? Yep. But I also know a subset of those proteins are close to my enzyme, like cutinase, for example, like a specific enzyme family that I can also use either directly protein language models or just the embeddings of it to tell me the direction that I want to go in. So the generation can be also multiple steps. And using the foundational protein language model, I may use only after that the embeddings of it as an input to my own generation model after that to come to the generation, because again, when we are looking at it, it's not just folding. It needs to be like Arzu said, we want thermal stability. We want to be folded in a certain shape, et cetera. So within the generation, we can also, not can, we are also using multiple layers and multiple models. Okay. And I love that you framed this from like a foundational model standpoint. A lot of these individual pieces, you're not building, they exist in the scientific community. You're leveraging just like an app developer is using OpenAI's AI API. You're using these foundational models and you're building the application layer on top. And your application is, help me find enzymes that will do this specific job. Exactly right. And even though we are calling it application, like you mentioned earlier, it has multiple pieces, right? Like it's the generation first and then activity function again. And then maybe there's another application and then the third application is prediction, et cetera, et cetera. Okay. So let's walk through this. In an application, I orchestrate LLM, not just LLM calls. If I'm building a RAG step, I'm first grabbing an embedding and then I'm storing it. And then I have, even before that, I'm chunking it. Like I've got all these steps. And then maybe in a LLM call, the LLM is retrieving from that. It's looking at what it gets back. Like maybe let's talk through your whole sequence. Like I'm not sure I have a clear handle on what your input is. And then what are the steps that are, I'm getting a clearer idea of what the steps that are happening. And then I'm also really curious about this from a success criteria guardrail. Like how do you know it's doing the right thing? And this is particularly interesting because it sounds like the goal of this whole AI workflow is help us make better bets about what to test in the lab. And so I can imagine mistakes cost you time and money. So maybe walk me through, what does that look like? Like when you're starting with it. Going backwards, you said the right thing. Actually, we were just talking about it in the sense that our AI platform's customer is our lab. And then our lab's customer is end-to-end process. So far, we've focused on delivering what our initial customer, our lab requires. The lab wants to test minimum amount of enzymes and it wants to get the highest performance metrics that we identified and can be multiple metrics, by the way. So we have done that actually in our platform version 1 that we call it. Let me describe that platform version 1 and then I will tell you the roadmap. Because the idea is really fast. In the first version, again, we didn't only use LLMs, but let's call it intelligence or generation, et cetera, piece, right? There was an human orchestrating these pieces, meaning Arzu did the literature database collection. Okay, what do we know from the literature, research, the nature show, et cetera. So we did that with a human. And then let me just, I wanna make sure as you go through this that I understand each piece. I imagine that research is like, what are enzymes that have certain properties? Where are we getting insight from? What's, you talked about enzyme families. If I'm thinking about this from an embedding space, what part of the room do we wanna be in? Exactly. From multiple different angles, specific to our function, specific to our polymer, specific to our family, the performance metrics they identified or what has been done in the literature. So that was step number one. Okay. And then again, I'm the orchestrator, so there's a human orchestrator. Yeah. And then we have these other modules that computationally we built, including very simplistic again, like three pieces, maybe generation, validation, and prediction, maybe let's say. And then we built those, but every each one of them, I orchestrated that workflow. Yeah. And then output went into the lab and then we tested it. And then there is now private data that we have a benefit of that we fed into this flow. Yeah. But more interesting thing with the latest agentic workflows and the power of these intelligence modules coming up, now we are putting all these together without human in the loop. Maybe except the wet lab. And in the future, we will also remove the wet lab too, but let's not get there right now. But currently, our platform can generate this research that Arzu did in the past. Let's assume we're gonna tackle another plastic target or we wanna improve a different key performance for the enzyme we have on hand. So now we have putting all these pieces together with an orchestrator, now also an intelligence, maybe a LLM model, et cetera, et cetera. So it is a huge for us because it is improving not only the final product, but also the development cycle for us, because AI is a development platform for us. Therefore, we can make it more faster or less computationally is also beneficial, but also of course the end product will be better than what two of us can do. So orchestration is really key. And we talked about multiple pieces, right? Like we said, data collection or literature, data accumulation or knowledge base. We talked about multiple different problem spaces. All of them think about as different agents. But we need an orchestrator. And then eventually the web lab, maybe that's the human in the loop piece for us, web lab data. But that also needs to come back into the orchestrator at one point and continue where it left off. I also can imagine even your problem statements and how you're constructing them and the constraints that you define and like each piece of your pipeline or orchestration might need different context about the problem space and that there'd be a lot of parallels with just what to give the models when and how to move that through the whole system. And models, and again, since we are moving towards orchestration of multiple agents, define handling these different pieces, the orchestrator needs to understand the full context, but also provide only the necessary or important context to each agent, for example. If you have a module, I have an agent who only takes care of the folding, for example, then it doesn't need to know anything else, but it really needs to know what are my, piece. And even in the prediction, like, there's a, you violated a constraint, you've got to go back to, maybe even back to the sequence. Is that now that you're moving to this orchestrator agent, is that kind of what it's, is it coordinating between the steps? 100%, but I also wanna correct something. Even though it looks like deterministic, nothing in this flow is deterministic. Yeah. You are following it still a statistical model, right? Compared to what humans used to be able to do in the past, it is pretty deterministic, but... I'm sorry, I meant deterministic in terms of, first, the sequence is generated, then the folding prediction happens, not that within the step it's deterministic. That's right. Like a lot of teams have gone through this with just even like web features. They start with a pipeline. Like, each step might be an LLM call, but it's a, it's not agentic, right? It's just a pipeline. And then as the models get better, teams are moving more towards, let's actually just have an agent and each step of the pipeline is a tool, and now the agent can decide, am I ready to go to the next step? The next step broke, I gotta go back to step one. And you get a little bit more of like non-determinism at the flow level. Exactly. And further than that, for example, we are simplifying some of these modules. Let's say we are calling about alpha fold folding. Yeah. But for folding problem, I can also have other tools as a signal. So agent, my folding agent, can use alpha fold, but also can get more other tools to use, get more signal. So that is also autonomy to the folding agent to do with the goal and maybe multiple tools and then get the orchestrator with the potential results or findings. Okay, so we get all the way through your process and out comes, I'm assuming the output is, here's a compelling enzyme experiment. Give me a sense of what has been the impact of having this. Are you running, are you able to run more wet lab experiments? Is the goal to run fewer and have a higher hit rate? What's been the impact of having an AI scientist on your team? The impact is, again, on the definitely the timelines and also the diversity and the design space we can get it. That is, I will make the, I'm only able to explain it with a comparison to the old school way of doing it. In old school way, generating diversity was a stepwise thing. When you engineer enzymes, you identified positions and the amino acid changes that make an impact. And then you go and combine them, but every time identifying takes a lot of time and then combining that takes another amount of time and resources. Now we bypass all of this and get different diversity of design in our hands to go and test. And that is like changing the paradigm in a way. And also the changing the timelines, because the amount of time and resource it takes for math is completely different than what it takes if you do it the old school, which is literally manual human predictions and suggestions and then go and test them in the lab. The other alternative is doing it really more like a gunshot that just without putting any rationality, just test many different versions without knowing where it will end up. And but then it costs a lot of money because you need, you need to test in physical world. So definitely the timelines and the diversity, the, what I mean diversity is the space we end up in. It is not like an incremental thing. We end up being in a completely different design space. I can imagine in the manual process, a lot of a scientist's like hypotheses about different enzymes is really influenced by the literature they have followed, their past experience, and that the diversity is probably pretty low just because of like human limits on what we can know, right? Yeah. The way to search for diversity was actually going to the nature. That is how we created diversity. Because we knew that we are limited in the number of sequences and diversity in our hands in the lab. But let's go to the nature. Nature has evolved many different organisms. Go and look what they have generated and copy from that. That is how experimentally we did to generate diversity. Now we have a tool to generate diversity in a much more quicker manner. And I can imagine like all that knowledge of what we know from nature is what's going into that protein-like design. Totally. That is what makes it possible to begin with. Yeah. But allows us to go to other places. And generally I like to present it this way that the design space is huge. Think of it this whole universe. And nature has evolved it to that much of that design space. Now we are able to explore all the other pieces that the nature hasn't evolved into. And that is for me mind-blowing, I would say. That's a really great point. Like just thinking about like at the human scale, thinking about taking inspiration from nature, that feels so big. But what you just said, I want to just highlight that. This space is even bigger. And the models have the capacity to explore this even bigger space. So like the space that felt overwhelming to us was a pin. That's very cool. That was very awe-inspiring. Yeah, because as said, our reference point is always nature in biology. Like you go to nature, you start from there, come to lab, and start engineering in a rational way. That is what we did. If you want to generate more diversity, you are going back to nature and explore different environments to create that diversity. But you, as said, you are still limited within a space there. And AI is bringing something else. And I'm really curious to see where things will go and end up. We are, I think, at the very early stages of all these developments. It's fascinating that this nature has benefited from millions, maybe billions of years of evolution. And yet it's still like a tiny percentage of the total search space. And like, I'm having this moment of like when you read, when you learn about the cosmos and you get a like sense of just how big it is and like you start to struggle to wrap your brain around it. It's a little bit of the same thing. We think about the diversity of living organisms and that already is huge. And now you're telling me, but it's just a tiny piece of what we could be exploring from an enzyme standpoint. That's amazing. I love that. You looked like you wanted to jump in earlier. I was going to say, build on that. I'm curious because you said like, you also talked to other product, for example, teams, et cetera. I'm sure the LLM hallucination is a problem. Yes. In some cases, it's not for us. And I was just thinking, for example, in some applications or some modules, I want it to hallucinate. Yeah. Because if I provide the existing research as a context, like she said, then I am also limiting what you guys just talked about, right? You want to explore that design space. So in some cases, the hallucination is that, you want to use, yeah, you want the creativity mode. Exactly. Yeah, this is great. Okay. Tell me a little bit about what's next. It sounds like you've got a pretty good system in place for predicting enzymes designs. I imagine you could iterate on that infinitely, but you alluded to there's more here. Like I said, the next one is, instead of thinking our customer, computational biology customer, AI customer, the enzyme design, making to the next one, which is the ultimate goal, having an end-to-end process that is feasible business-wise. So currently, our startup, we are scaling up this process to 5,000 ton a year, like a demo plant in California. So we are developing this orchestrator model, agentic model, not only enzyme design, but we are introducing like a process agent, for example, which understands our process and provides inputs or develops maybe new process parameters. So that's the next step, which is introducing this process angle to it, but also make everything more autonomous and loop, and ideally, not right away, but we also want to get to go to wet lab testing at one point. If predictions are strong enough, right, then I will reduce that. Our predictions are strong enough, we will reduce that significantly. So, if I remember right from your application, you're partnering with a bottle lab in California. That's correct. So that's how we gonna scale up our process in California to a demo plant plant. Okay. And the goal there from a process standpoint is, it's great that we have this like prediction to get to an enzyme, but there's the back half of the problem of we've got our reactor, we've got our enzymes working. How does this fit in the bottling plants ecosystem? Exactly. Even within our process, after we break down those monomers, we need to separate them. So that's another process. When we are doing the separation, we are realizing there are some effects of enzyme in the downstream separation process. So there are new potential process or enzyme parameters that we can focus our AI to target, which are not a typical enzyme design, for example, which is more specific for our process. Okay. I can also imagine if you get to a point where you can really scale this, you almost could do like custom enzyme design for a customer, right? Like a customer's product is specific, designed in a specific way. You're designing an enzyme specifically for that product. And then this process agent can also learn about specifically how that product is recycled. And you're almost getting to bespoke enzyme design and process integration because the AI is driving all of it. Is that part of your vision? That's absolutely correct because we talked about the foundational level of our application and that depends on the abstraction level, like how much we abstract or not. Our