 Welcome, welcome. Hello, Kronan. Good to have you here. Hi, Nicos. Good to be here. Good to see you again. This is Nicos Manuselis. I have the pleasure and honor to be hosting a series of fireside chats on AI and food safety. I think about the most relaxed conversations, the type of conversation that I would have with you over a glass of wine, over a cup of coffee or tea, trying to demystify, uncover what is behind artificial intelligence, AI, and its applications in food safety and food risk prevention. I have to admit that you, Kronan, was, I think, number one in my list. The top number one person that I wanted to invite in this conversation. So, thank you for being here. Kronan McNamara is the founder CEO of CREM Global, extremely widely known as, I will use my words and then you will give me the proper intro, AI modellers that protect everyone's health. And what I really like about your work, Kronan, at CREM is that it seems that you're working on local problems that have global impact or global problems that have local impact. It's amazing, at least in the way that I understand it. So, let's start by who you are and what the company is doing. Okay. Thanks, Nicos. And thanks for pronouncing the company's name correctly, CREM Global. So, my background is physics, actually, and maths and computing. So, I got involved in food safety research during kind of post master's postdoc research in Trinity College. And I found it interesting to kind of apply the mathematical modeling and Monte Carlo type simulation into a new area, which was kind of interesting with lots of data, lots of complexity and interesting challenges. So, we got involved in food data science, food science modeling back in the early 2000s. And I founded the company CREM Global in 2005. And we got started working with some industry data and regulators and industry on both sides, looking at different aspects of food intake, food exposure, and then moving into more machine learning and predictive modeling as we went along. So, it'd be interesting to talk to you about that a little bit today. So, when you started, it was something like 20 years ago? Yeah, nearly. Okay. It was all about mathematical modeling and predictive modeling. And how did the name come? Oh, that's the most frequently asked question. So, I'm happy to explain it today. So, when we were in Trinity College, there was a research project and it was named by Professor Mike Gibney, who sadly passed away last week, actually. So, very sad news on that front. He was an inspiration and great at bringing consortia together to do research. And he was a professor of public health and nutrition in Trinity College at the time. And he came up with this project called the Center for Research in Exposure Modeling Estimates, which was an acronym for CREM. So, we worked on that project in Trinity College and it followed on from an EU project called Monte Carlo. And as we went along and started to engage with industry and government, they got to know the name CREM. So, when we formed the company, we thought it would be sensible to keep the name CREM because it would already have a little bit of recognition out there, but we dropped the acronym. So, that's a most frequently asked question. We tried to come up with new acronyms, but eventually we just left it and just called it CREM Global. And you started working on exposure modeling problems. What do you do today? Well, yeah, exposure modeling is still a big part of the work we do. And we work in food safety, agriculture, chemicals in cosmetics. Those are probably our three big domains. I always liked building platforms and products. So, we invested a lot of time and energy in developing our own cloud-based platform, which is a secure way of aggregating and gathering data, which we can talk about a little bit more. So, we help industry collaborate by sharing data in a secure and anonymized way. And then that data can be visualized or modeled using predictive analytics or machine learning type methods. And at the same time, it can also be put through more, I suppose, mathematical models that we've created over the years, which really just look at exposure. And that can be combining consumers' habits and practice data, which we get from various sources, either public or through purchasing data. And then combining that with ingredient and formulation data. So, then it can kind of look at a very detailed level of what ingredients are being consumed or being used in cosmetics or being consumed in food and adding all that up. So, it's a simple enough model of adding up lots of stuff. But the complexity comes in the uncertainties. So, we use what's called Monte Carlo simulation to model all of the uncertainties. So, any number or input in the model can be represented by, you know, probabilistic kind of range or distribution to help simulate that uncertainty and variability. And then give good predictions of exposure or intake with looking at the high consumers or the average consumer and creating those kind of that type of information for government and for industry. Okay, now you will help me understand this a little bit better. Because I hear you describing different things. I will park for a while the conversation around data sharing and private data sharing, which I find fascinating and it's essential. It's one of the topics that I really wanted to talk about with you. But you said that you started working on mathematical modules and then in a very sophisticated way you incorporated in this mathematical modeling, also the parts that have uncertainty. Yeah, exactly. Parts that cannot be determined in a certain way. So, at which point of this journey did AI come to play? And how? Good question. And a bit later in the journey, because our core offering when we started the company was were these types of exposure, mathematical models built from first principles, using scientific knowledge around intakes and different factors that are important like absorption factors and things like that, that we can model. Because you're never sure of those numbers, you have to use a probabilistic approach. And these mathematical methods I was using in college, I actually did a bit of work in the financial industry as well, valuing complex derivative options, again, using probabilistic methods where you don't know the future really, you have to only know the present, but you know, the future will diverge in various ways within known limits, I suppose, are reasonably well known limits. The financial industry didn't always get that correct, as we know. And they overestimated their confidence in those kind of limits and things that can go wrong. But yeah, exactly. So any uncertainty can be, you can model it using these Monte Carlo methods. And that's something I found very interesting in college. And I've enjoyed implying that into food safety risk assessments or exposure assessments these days. And what you're saying is that the starting point is a number of set first principles as you call them, which are the input factors or variables that might affect the outcome of the model. And that at least in traditional mathematical modeling in my basic understanding, they were given as something that is for granted by the scientists in its field. They say, guys, we know from our research that these factors correlate with an outcome that you should expect. Yeah, exactly. And when did AI come to play? And how did this change your perception of the problems? Yeah, so that's exactly right. So these are from scientific first principles, we would collaborate with toxicologists and with nutritionists to understand the key issues. And all of these parameters do affect the intake and the exposure estimates. But the question is how much do they impact it? What are the key drivers and how confident can we know about that? So those are those are one category of models you can use, which are scientific models or mathematical models from first principles as you call them, which is correct. And then, you know, I've always had an eye on the machine learning aspects as well, always fascinated by that approach, and never had a cause to apply them in those exposure type projects, because, you know, these first principle models were working very well, and they're well understood, kind of more transparent, I guess, as well than an AI model, which kind of comes up with predictions without necessarily any scientist saying, you know, programming in the different relationships. But as we went through a number of different projects where we were gathering new data and asking different questions, essentially, not just exposure questions, but other types of questions like, for example, food fraud prediction, or outbreaks of a pathogen in a manufacturing environment or in a agricultural, you know, in region, you know, these are different questions and harder to model from first principles, because we don't know the first principles a lot of the time. And therefore, these machine learning methods are really interesting to use to just discover correlations and categorizations of different things within the data that white would not be obvious, you know, to humans, but just by looking at visualization, for example. So different challenges require different models, I guess. And that's where we when we had different challenges, we moved into the more machine learning AI type models. And there will be the types of problem short challenges that require the machine to build the model as it goes. And as it gets new data to generate this mathematical model on its own. Do I get it right? Yeah, exactly. So even just training it on a lot of historic data, obviously, you have once you get there, you have a model that you've kind of trained and tested and you're reasonably happy with, or historic data or more new data comes in, you can retrain the model and update it if necessary. Yeah. Or keep testing it as you go and see if you need to retrain the model. Yeah. You have two scenarios, one that still works extremely well when you apply the traditional mathematical modeling approaches and one, but where you see machine learning approaches working best as simple examples to understand the types of challenges that fit better to each one of the approaches. Well, exactly. Like being, I suppose, a physicist as a background, you know, some of the physics we know the maths of we can just use models. And similarly, in some of our intake or exposure modeling problems, for example, we might be asked to look at the exposure of a population to a pesticide, we have a lot of data on food consumption, we have data on agricultural conversions from food to the raw commodities, we have data on the pesticide monitoring program. So these are all three different distinct data sets that were never designed to be used as one one calculation in the model. But by using like first principles, we eat food, we measure that we know the person's information about the person, maybe their age, their body weight, and we know the amount of food they're eating from these kind of studies that come from government or, for example, the NHANES database, we know the information of the commodities we know the pesticides really just have to, you know, we really just have to kind of combine those in a sensible way, you know, and that will create a first principles model that works. You could train an AI model to do that eventually, if you did enough traditional studies and knew the answer and said therefore, you know, new pesticides has been developed here some data on it, the AI could learn those patterns, but it's not really necessary because the first principle approach works well and is explainable. And but in other questions, then, where you need to go into AI and machine learning, for example, are much more I suppose, they're sort of different questions in science where we don't know the cause of everything and it could be human factors that apply to things like, for example, food fraud is one interesting one, you know, when climate changes, when prices of commodities change, when, you know, different political things happen around the world, you know, this stuff as well, you know, you're looking at these questions yourself, you know, there's no real first principle model that says, well, when the political instability goes up, therefore, the risk of fraud goes this way, you know, there's no physical reason for that, you know, scientific kind of principle you could apply there. So therefore, you have to measure that using data and, and, and, you know, in the old days, we use statistics and just said, well, this looks like there's a correlation there. And that's really what machine learning is doing, right? It's correlation testing on steroids, all the time changing parameters and trying to find fitting of these many different parameters. And that's the way I see machine learning is really just automated statistics on steroids, just finding patterns and optimizing things. And therefore, when you finish that process, you end up with a model that can actually, on new data, make a prediction, right? So the final outcome or the final IP, what you're generating, and is your, is the company's knowledge IP, it's the model percent, right? It's a trained model percent in both scenarios, right? Yes, the IP of the of the company Kramglobal is the data and the model, I guess, or whatever curated data, let's say, our access to data is a really important part. And then the model itself, I guess, is part of the, you know, the trained model and how you go about that is important. We also have the, you know, we also think of our cloud infrastructure as IP as well, I suppose, because we've engineered a system that can host models, host visualizations, and gather and help organizations manage and gather data in a collaborative way. So different aspects of that, I suppose, would be considered IP. So what I hear you describing is that technology per se, which we would expect, being from the technology component world, is an IP on its own data. Let's talk about data, but it can be our data, or it can be data that is provided by third parties. But the model per se, and the trained model that deliver reliable predictions is something that is being developed by a company like Kramglobal. Yeah, let's talk about data. Where is the data that you're working with coming from? Yeah, there's different sources of data. And there are public data sets that we can access and governments publish these. For example, the NHANES, the CDC in the USA, would publish an NHANES database, which is very rich and has lots of information in it and is useful. Other data sets like the Agricultural Research Service at the USDA and other PESA side programs, monitoring programs, and all of these data sets can be published. And other European and Asian and South American countries have similar data sets. So that I always like to kind of start from the public data and see what's there and what you can access. Sometimes you have challenges around accessing the data and seeing what permissions are associated with that data and trying to negotiate that process. So that's a good start. But the thing is everybody has that data, right? So for a company, it's not necessarily a competitive advantage to have that data. You can spend time and effort curating it and making it useful, but there's some value in that. But then the real interesting part comes when you have industry data or somehow working with industry clients or even government clients to gather private data, be that around the products that they're creating, the formulations or the monitoring programs that they're participating in, in private industry as an individual entity. And there's certain value in their own data. But when they start to share that data as an industry group, the real value emerges because you get a bigger picture of what's going on in the environment, in their sector. And you get enough data to train a model. As we know, machine learning models are hungry for data. So the more data you have, the better you can train a model. So it's great when organizations start to pool data that can be used for training machine learning models. So one source of data is coming from the public sector world. And this is not, although it is something that everyone can access or try to access and process and use. It's not as easy as it originally sounds. Exactly. And then the second source that I hear you describing is a customer's data. An organization is coming and is sharing private data with you because they want to build something to address one of their use cases. Yeah. But you do highlight, again, it's, I think the second or third time that you, you highlight this in our conversation, the value of getting more than one organization pooling together their private data. How open are they in doing something like this? Yeah, that's the challenge for many reasons. And it's always a slow start to these kind of initiatives. So some industries have done this really well. And actually at the SOT conference coming up, we're going to do a little joint presentation with RIFM, the Research Industry for Fragrance Materials, on how they got it right. How they've done this for many years and really shown a very successful case study of industry sharing data. Now, these are fragrance formulations, which are very, very highly secretive. These are the secret sauce of perfume, fancy brands and perfume, that they don't want people to know what, how they formulate those mixtures to make these expensive perfumes, and then they sell those onto other product manufacturers like shower gels and shampoos and body lotions to use those same fragrances. So very proprietary information. They would never share that data with their customer, but RIFM, through good science and good trust building over the years, convinced them to start sharing that data with RIFM, the Research Institute for Fragrance Materials. So now they have done, they've gathered data of over 2,800 fragrance ingredients, they've got the most comprehensive database of formulations used in the whole cosmetics and personal care sector. When the government has a question, they go to RIFM, when industry has a question, they go to RIFM, and we work with RIFM to help them gather that data, and we've put together the model of exposure and risk that's used then to set safe limits for all of those fragrance ingredients in collaboration with RIFM. So it's great when you see success stories like that and you can use it to motivate other sectors to try and do something similar, but they're always slow to get started. But when they do get started, as I think you know, in other projects, they can really build up a strong trust and collaborative community there, that actually tends to really work well together in the medium term. And what I hear you describing again is that the real value is in creating a resource that is of value and supports and delivers value to everyone involved, all the stakeholders that are involved in sharing data. In your experience, what is the aha moment where they say, OK, now we get it? Yeah, hopefully as you can deliver that aha moment early in the process so that they kind of get small wins or even early wins. They call it low hanging fruit, sometimes at the initial stages of a project, because if they don't see it and they're putting a lot of effort into data collection, and usually there is some effort involved, internal to their internal infrastructure, internal administration and trying to organize this data. And also then they're worried about sharing the data in the first instance. So if they don't see results reasonably early, they can really lose momentum. So it could be from the first visualization that you do on the aggregated data and they're comparing their own industry or sorry, their own company with the aggregate. That's a win. You know, oh, we're doing well here. We're not doing so well here. We're better than average on this. We're not we're worse than average on that. So that's that can be just a simple visualization. And sometimes that's impactful, you know, dashboard that can they can interrogate and see. And that's that's a nice way to start. So building up a nice bit of data visualizing it. And then as you have enough data, you can start to start to try to train a model. But I wouldn't go straight to trying to try to train a model. I'd go straight to a visualization initially just to give them some value quickly because training a model, you know, they'll be like there's nothing there for them to see initially unless you can visualize the model and the results of the model. And will they really trust the model? Not really. Probably not initially. So that that's a time it takes time to trust the model. But they can trust seeing the data and transparency of the data to us to as much of a level as you can do in the aggregate worlds. You can't give away, you know, information that could be giving away private information of one of the participants. So you have to be careful with the aggregate visualizations. But then you can give them their own private visualization of their data, which is is more detailed. And that's that's a nice way to start these kind of data projects. So even pulling together the data, the aggregate ones or even a better version, a more curated and high quality version of their own internal data and developing some initial visualizations that can help them understand what they see in their own data and benchmark themselves. Yeah. Against others is a strong argument. Yeah, this is what you said. And you also said, as far as the model is concerned, and I thought that you would say when they see the results of the model, they get crazy and the bite, you said, no, it's difficult for them to trust the model. Why is it difficult for them to trust the model? What do you think? Well, there's a few reasons, I think one is they're already experts in this business they've been in for many decades, right? And they know they kind of know what's going on anyway, you know, in terms of food safety. Other industries, maybe there's, you know, there may be more trusting because they don't have the intuition, maybe food fraud is one of those markets where they might trust a model better because they don't really know what's going on. But if you're a farmer and you've been growing product for decades and you know the risks and you know what tends to happen when it rains or if there's a storm or other things, you know, if you're predicting those things they'll be already saying, well, you know, I know better than this model, you know, and it's going to take them a long time until the model proves itself worthy of their trust, right? So that's what I think I spoke is one of the reasons that they're already experts, right? And they and if they're clever, they're going to be a bit skeptical and they should be skeptical of a new technology until it proves itself. Do you get do you get this kind of skepticism? Well, yeah, definitely. I think, you know, oh, that works great. Like, you know, you can tell the client, oh, you've tested the model on like thousands of records and you're getting 95 percent, whatever confidence and accuracy. And they're and then they, you know, they're like great, but I'm going to test it out for a few months before I trust them. And you're like, well, that's what we just did using your historic data, but they don't necessarily, they want to test it themselves as well, right? And what you're describing is creating this trust and creating this environment where people are sharing organizations, it's even more difficult, are sharing data with with each other in a controlled environment in a group that has agreed on data sharing principles. Yeah, what would what would it take? I wonder to continue this journey and making such insights available for all? How far are we from a future where this valuable knowledge that you see and generate, especially when you're looking at predictions, where we can make it something useful for the rest of the world? What do you think? Well, yeah, that's a really interesting part. And in some of the projects that's already possible, that they're not saying that only people who participate in the data sharing can use the model. Actually, some of them are saying anyone can use the model. Now there is a cost involved in using the model just because it maintaining the model and all of that. But some of the projects we're involved in already have gone that step to say this can be used by government, this can be used by other sectors of other companies that haven't necessarily been a part of the project in the sharing of the data. So definitely that's possible. It depends on, I suppose, how sensitive they are to the data. And maybe in some of the other projects, we could look at sharing a subset or a more even more anonymized version of the data to give people some insights. But there's always a risk with that where some NGO or other organization with an agenda will just pick up some one or two points in the data that look bad and, you know, make a story out of it and cause a lot of trouble for an industry. So why would they want to sort of take that risk where they don't have to? What's the benefit to them of doing that? So there's a challenge, obviously, there to try and overcome. But maybe I think I really like the spirit of scientific sharing of knowledge. So writing maybe scientific papers or things like that around around the results that have come from the data could be very useful to share the knowledge and learnings that could be translated to other regions or other industries, right? This makes sense. Yeah. And do you have a problem or a challenge or an area that you haven't touched yet, but you are eager to look at? Do you have a dream problem that you'd like to attack? That's a good question. I hadn't thought about it, but yeah, lots of ideas. I think I think nutrition is an interesting area and causes a lot of health consequences, probably even more than some of the food safety things that are far more high profile. So malnutrition, either over nutrition or under nutrition, I think is a very interesting scientific area that could have huge impact on health and people's well-being. And I'd love to do more in that. We did a study many years ago on personalized nutrition called Food for Me and recently on a new programme that took a twin study in Stanford and looked at the genetic impacts and things like that. It was it was interesting to see the study that they did with twins, right? So the genetics and the impact of genetics on people's diet and on their lifestyle as well as other factors. I thought it was fascinating. So I'd love to do more in that space. So this would be the space and the the area where you would like to get more data. Yeah, yeah, have access to the data and to build the models and try to see what you can predict in terms of expected outcomes. Yeah, real science into it. Like there's better tools now like everybody's wearing some kind of fitness tracker, like a watch. I've got the Apple Watch. It's got great data, monitoring my health, all these different metrics. You can get, you know, more genetic data quite easily these days, cheaper microbiome information, all this stuff. So I suppose it's just an interesting challenge with all that complexity and lack of understanding of the interaction between these things of how to live a healthy life and optimise your lifestyle under under the various conditions of people. So if we have this conversation in five years or ten years from now, what would you like to have done, have achieved and be very, very proud of? I think these I'm not sure if we'll get into that that area in the next five or ten years. Maybe we will. But I think we'll continue on our journey around all the other aspects of food safety and chemical exposure and nutrition. We are doing nutrition work, but that's a very ambitious grand challenge, right? And maybe we will try and set up something like that, maybe as an EU project or as an even bigger worldwide project. I'd love to do an EU project someday where I kind of came up with like a grand challenge like that and could have the time and space to try and really work on something like that for five years. I find a lot of the EU research projects are, you know, there's some kind of science behind them, but nearly it's already known at the start what's going to happen at the end. And you do your bit and it's a bit formulaic sometimes. But I'd love to do a more ambitious research project with some good resources behind it if I had time and the resources to do it. Yeah. So you would like to have spent five years already or more devoted on shortly one of the grand challenges and having the time and resources to do it. I really wish that you tell me such a story when we have this conversation. Yeah, maybe we will make maybe I'll have to wait till I retire to do that. Okay. So if you would like our audience to keep one key message from our conversation today, what would that be? On the topic of AI really are generally or whatever else we thought stop on today. Yeah. Well, I always try to emphasize like really good scientific thinking and being critically critically evaluating, you know, data and using strong mathematical methods, you know, so I think that'll always be important, you know, we see these days, the emergence of AI systems like the chat GBT and ML, LLM models, large language models that are becoming very, very, very powerful. And people worry about, well, where does that leave us in the work we have to do? And for now anyway, we still need to be very good critical thinkers in order to interact and evaluate the output of these models, like I was saying earlier, you know, if you provide a predictive model, even to your customer, they're going to be skeptical and they're going to think about it critically and see if it makes sense and take time to trust it. So I think being able to evaluate things like that, and being good scientists, good, good mathematicians and will be really important in every industry really going forward. So I think that's a good takeaway for your takeaway. So let's stay good, critical thinkers and evaluate what such technologies can offer with critical and scientific. And you said with a mathematical eye as well. Yeah, that's a powerful message. Yeah, thank you so much. Mathematics is probably the language of science really, isn't it? At the basics, the language of science. That's a punchline that I will keep and use to promote our conversation. You can use that as the headlight for this. Thank you so much. It was a pleasure to have you here with me. We've talked about amazing things and we covered lots of different aspects. Thanks for the time and the energy and the openness. Oh, pleasure. Thank you, Nick. Us. It's been a really interesting shot very enjoyable afternoon.