 A very warm welcome to our webinar, the titles microtoxin risks in the food and feed supply chain. What I'm going to do, I'm just going to wait about one more minute just to allow people to join. And then I will continue to give a little bit of background information to what it is we're hoping to achieve today just in terms of the information that you're going to receive. But also, very importantly, the information that you are going to supply to us, the panel as well. So we'll start very, very soon. Okay, I'm going to start in about 30 seconds because we do have one hour. Well, I think this one hour is going to pass really, really very, very quickly because we do have a lot of information that we want to share with you. So we want to get the feedback information that you're going to be asked in terms of some polls and below. So really importantly, think about the questions that are coming to your mind as we go through all of this different information. And then please by all means put them into the Q&A box which is at the bottom of your screen. I'll do introductions to our panel members in a moment, but I think first of all, what I would like to do just to set the scene a little bit about my toxins. Now clearly the fact that you have decided to join our webinar today means that my toxins is a subject that you have an interest in. Well, I wonder, and you now realize that the most recent data, and I was part of the group that put this together now established that depending on which crop it is, somewhere between 60% and 80% of all crops now contain mycotoxins. So basically think about virtually every material that you have to deal with, which is plant based is a problem for you in terms of mycotoxins. It's not just about food safety risk, we're going to talk about some of the risks and that that are coming up, but also it can have a massive financial impact on your business. It can have a massive in terms of negative performance on both the welfare and the performance of livestock. We're getting more and more information and worrying information about actually the environmental impact on a carbon footprint associated with mycotoxins. We now are very, very concerned about food waste and reducing food waste, but actually mycotoxins is one of the really big contributors in the world to mycotoxins out of food waste. So lots and lots of things that we want to look at. So really our focus for today at the webinar, be it morning, evening or afternoon for you, is thinking about new ways of moving from being reactive to a particular problem, the mycotoxin problem to being proactive. And that is really a very big challenge. So we're going to talk about the risks in terms of what mycotoxin bring to and how forecasting can actually bring those risks to your desktop really very quickly and in a very accurate way, and then allows you to think about how you can deal with it. So that's what our panel is going to discuss and debate today. So mitigation of mycotoxins. I mean there are lots and lots of different ways that businesses try to deal with it at the moment, and generally we divide it into pre-harvest into post-harvest. And some of it is around getting good sampling, good analytical methods, the application of good agricultural practice, and also there are really issues that are coming to the fore, not even about regulatory frameworks, but about rapidly changing regulatory frameworks, which actually is an additional risk that we all have to think about. So there are different ways that you can think about detoxifying mycotoxins, but really what we want to think about is this predictive analytics side of things. How can you predict what mycotoxins may or not be in the particular crops that you're working with. So this is the time to introduce our panel. I'll start off with myself because I've been talking a little bit so far. And my name is Chris Elliott, I'm Professor of Food Safety at Queen's University Belfast and Professor of Food Security at Tamizat University in Thailand. And one of our big projects is mycotoxins, particularly associated with livestock. And I'm really delighted to be joined with some really industry experts. I'm the academic who just sits and writes things. These are the experts who actually do things and have to deal with lots of problems. So the first is Maria and Willis Garry, who is the ex-corporate R&D and VP and Chief Scientific Officer from Mars. And we're also joined by Joe Tierney, who is regulator and food safety department lead at Terlan previously known as Glambia. And our fourth and final panel member is Gianna Stoictis, who is Chief Technology Officer and partner of Agrino, who is hosting our webinar today. So many thanks for Agrino for bringing all of this together. So I think now it's my job to stop talking and start listening because I'm going to pass you over to Giannis now. Thank you so much, Chris. And thank you for this very interesting intro. I'm very happy to be in this webinar, in a webinar that deals with a very specific risk and I'm looking forward to getting your feedback about the risk of mycotoxins in the supply chain and the global specific supply chains that you are responsible for. So I would like also, Maria and Joe, to make a short statement and then I will move with a short intro about the work that we are doing. So, Maria. Thank you, Giannis. It's great to be here today. This is such an important topic that touches on every single ingredient, pretty much plant ingredient and food that we consume as well as feed. What I remember is feeding people and feeding animals. So it's really important to use the emerging and fast growing AI tools to do a better job with predicting prescribing solutions and mitigating apoptoxin issues. Mycotoxin and apoptoxin, which is one of the most it is very common in food to be able to predict them. Thank you for the invitation. Hello, all, and thank you also for the invite. I'm delighted to be here today. Certainly from an industry perspective, it's very encouraging this dual work between academia and obviously Professor Elliott and the team here at Agrinol. Just to look at predictive modeling related to mycotoxins and I'd echo what Maria said, there's a lot of these huge benefits from an industry perspective around predicting before we receive grain for for processing. So again, looking at the metadata and the predictive model and in more detail and seeing how we can use that in our risk profiling. So I'm very encouraged by it all and happy to be involved in this discussion. Thank you so much. So before going to the to a short intro about the work that we are doing. We have a very short poll about which is your how you rank mycotoxins risk in your supply chain compared to other risks that you have so if you if you want to give us a short answer to that. Thank you so much for that it seems that half of the participants that we have today with us and thank you very much for joining this webinar. We are ranking mycotoxins risk as the as a high risk compared to other risks. So this is very important. And medium and low are equally split it to almost to 25%. So thank you. Thank you so much for that. I will make a very short. I will be sharing the results. I will make a very short intro about the work that we are doing in on building tailored AI models for specific supply chains. So, first of all, agronome we are the data analytics company that uses AI to predict food safety risks. And our promise is that we deliver to the companies that we are working with a reliable risk forecasting for critical raw materials and ingredients by thoroughly tested and I highly accurate AI models. I think we are continuously measuring our promise, and by using accuracy metrics like the accuracy, the percentage of the accuracy of each deployed model, and also use application oriented metrics like the percentage of the recalls that have been early highlighted as an emerging risk using our models. So, in this way we are trying to measure continuously the promise that we are delivering and the services that we are the AI services that we are delivering to the companies. The main concern that we hear in the food industry, every time that we are talking to experts like you is if we can really predict important risks in the supply chain using AI models and today, we will try to answer this question for the case of my potoxins but we are, we are already applying and trying and deploying AI models from 2021 that are trying to forecast the risk trends for several ingredients raw materials. And the models and we are providing these models through dashboards in our food archive risk intelligence platform. The way the approach that we are following to build the AI models is that we start by creating by identifying which are the very important factors that we need to use in our models. We start by creating the tailor made data sets for these models and specifically for each ingredient for each region. And then based on the, the business question that we are, we need to answer we select the appropriate machine learning approach. We need to do if we need to predict to classify something, then classification method is a good, a good option. If we need to forecast the risk level for instance then a time series forecasting method could be a good option so we select very carefully which approach we will follow. And after selecting the right approach, we are training and testing the prediction model, and we are trying to refine the parameters of each model to achieve the best performing the best performance in terms of accuracy but also in terms of other criteria that has to do with the application of the specific model. So this is the way that we are developing such models, but I will now hand over back to Chris, because I would like very much to hear from the experts, which are the key questions in the industry that we need to answer, regarding the Yeah, many, many thanks for that, Janice. And this is the time to start to think about the key questions that you would like to ask. And what we have done is we really start to think about what we think the key questions are. And for our panel members, Maria and Joe, they had some time to think about this reflect upon this and then feedback some of their thoughts. So maybe Maria, what we'll start with you just in terms of using AI to build stronger testing and de-risking strategies. What do you think are the most important points that you would like to make? Thank you, Chris. Ultimately, this is about our ability to make accurate forecasts for the levels of microtoxins and the types of microtoxins for specific ingredients, food and feed, that we are sourcing in certain regions and we consume in the same regions or other regions because here we're talking about the global trade. So here are some important considerations in my mind. And the first one is, do the companies today put the appropriate efforts to develop not just any predictive model for microtoxin but actually sophisticated models because we're talking here about a very big, widely spread phenomenon. The second question is, which factors should we take into account? And this is a very complex issue. So we should consider a multitude of factors like climate and weather, like cropping systems, seeds, agricultural practices, pesticides, the use of water and so forth. Also a number of geopolitical factors for a global trade and actually very complex supply chains. And as Professor Elliot said earlier on, the ever-changing regulatory environment, because microtoxins are not regulated in all countries and they're not regulated in the same way. Of course important, it's important to consider the testing and surveillance programs that exist within companies but also at industry level and how you bring this all together. It's fundamental to our ability to test and survey effectively is our lab testing capabilities and how we set up a risk-based laboratory testing approach and with the associated programs. And the last and not least is, how are we going to use these AI models within organizations and the gross industry to optimize the outcomes for the investment and how are we going to have the best return on the budgetary investment. And this is people and money so that we have the ability to forecast and then we have the ability to put in place prevention and mitigation measures. Maria, many thanks for that. I think what you did was you captured a lot of the big issues very, very well, but also the fact that what companies need are information about very specific risks as well. So thank you for that. Joe, we'll pass over to you now with your chance to kind of reflect on those questions that we posed to you earlier. Yes, so I suppose undoubtedly from a mycotoxin risk perspective, any type of prediction that is again robust and accurate in its construction is very welcome. I suppose one of the items around the how do we use the data and AI to inform processors in advance of risk of mycotoxin. You could argue that a red amber green or a simple rag status of the mycotoxins for the export markets is key. And I see that's being part of the offering there in development. So I think that's certainly to be encouraged. And again, I suppose I've listened to a few year talks previously, Chris, but the metadata supporting conditions around how mycotoxins for me the pre harvest post harvest. All of that is significant and obviously understanding the metadata going into the model to make sure that it's robust and delivers some early warning I think is point number one. Point number two is I suppose the report sampling. And again, I know this has been flagged several times but just the standardization of how you take a sample and ensuring the again the robustness of that program, such that it's repeatable. And I suppose all results regardless of where they're coming from can there's visibility around it and it can be relied upon. I think Maria touched on this around the testing side, certainly the standardization of the analytical methods, alternatives, it could be lateral flow or rapid method of some sort, or whatever it might be to measure the target mycotoxin of interest. And obviously keeping an eye also on the maximum residue levels as they are being issued. And again, compound feed is an example in process regulatory limits around it. And again, always there's the toxin in one let's say for milk is an example or other mycotoxins the dioxin of Eleanor. Now as an example again all of these will have proposed MRL that we need to be aware of and obviously the analytical method should match at least what's expected from a maximum residue limits perspective. And then obviously emerging mycotoxins, I think, keeping up to date with the science and obviously linking in with academia on the same and obviously being part of the industry groups to understand what's new and its impact on both the food and animal feed side of it. And then the last number four, which is learnings from the previous mycotoxin risk prediction models and I know there was a few that we discussed recently around previous work and literature published, and maybe some learnings around it. And staying close to industry, maybe opportunity to share knowledge, maybe even results and looking external and combining and pooling results and data and looking carefully at the metadata used to construct the predictive model. So they're the items I suppose of interest from our end. I mean that that is really great and that that's such a wonderful build on the information that Maria gave us. And you know again, often I say, you know, because you talked about metadata, and I say, good data in means really good data out, and bad data in means rubbish data and I think this is one of the things that Janis is going to talk about just in terms of how data is handled. And I think the example that you gave Joe just about, we are really in the eye of the storm about changing regulations and compound feed. That's where, you know, I think so many companies need help and support. So Janis, over to you to describe us to us, you know, how does this all work. And thank you, Joe and Maria for sharing this very interesting and very important points. And especially the questions that is very, very important to answer when you are in the industry. I will, my presentation will have two parts in the first part I will focus more on the design principles that we need to follow when we are trying to develop an AI model that will be applied in the practice. And I will comment and provide some information about many of the things that Maria and Joe touched. And the second part of the presentation, I will show you how such a dashboard can look like which are the operations that it can provide. And here in the Q&A, but also after this webinar, we would love to get some feedback about that. So, listening to the questions, but also I try to bring in my mind what we hear in the industry about the business problem. And the business problem that we hear has to do with the investment with the large investment that the companies are making every year for testing for microtoxins and how difficult is the choice between the frequency and the cost. And we hear them, we hear the experts saying that they wish that they had a software system that could combine and analyze all the different data that we are discussing here today that exists out there, but also inside the companies to help them to predict and prevent food safety incidents that have to do with microtoxins, before actually they happen. And the specific critical business decision that we would like to help to support is how to set up the optimal monitoring program for microtoxins, for the specific microtoxins, that will be risk based and will focus on the high risk batches that I'm getting in my supply chain. So, starting with this, having in mind this problem and this decision that we would like to support. I will share some thoughts, first about the design principles that we need to follow if we want to build to deploy an AI model that will have practical value. And then I will show you how this model and how the dashboard powered by such model can look like. So I would like to answer with this slide I would like to answer to the very important question that Joe made about the previous works and the previous studies and the work that have been done in predicting microtoxins so far. And it is true that during the last 20 years, we have several very important results reported in the literature about predicting specific microtoxin risks. The types of the models that have been developed are mainly for the mechanistic, the empirical, the hybrid and the machine learning. And here I would like mostly to emphasize three things. The first thing is that these models are highly sophisticated, and they all the models seems to have a good performance, according to the reported accuracy, and they are touching and they are applied in a very specific supply chains for a very specific region. The very good news is that we have already different types of models that has been have been successfully applied in real world problems. So that I would like to add that during the last years we have also machine learning models, either using simple logistic regression approach or Bayesian networks or even deep learning models. So, they are constantly new methods that are used in order to build predictive model for microtoxins. The second thing that I would like to highlight from the study of the literature is that the reported accuracy is good, it seems that these models are working well under specific conditions, and with specific limitations, but they are working well. And the third thing that I would like to highlight from the literature is that, of course, as it always happens, there are limitations. I would say that the limitations are, we have two types of limitations. The one type of limitation is that many of these types of models require substantial amount of input data, specific data, data that are provided information about the local conditions that a crop was produced. So this is something very important to keep in mind that without substantial amount of data, these kind of models cannot perform very well. The second thing, the second important limitation is that even in the case of models that can be built, even if we have data gaps like the case of Bayesian network models, the update of such models can be challenging and can be time consuming. And one comment about the limitation of the deep learning models which can be, can model complex patterns and they are very promising is that, again, it's very important to have large amount of homogenous data and computational resources in order to build a successful model. So let's keep in mind these three things as highlights from the literature, sorry. The second very important thing, the second very important principle, and Maria mentioned that, is that it's very important to take into account the risk drivers that may increase the likelihood of microtoxins in food and feed crops when we design such a model. And you already mentioned several factors, so I will not repeat them, weather and climate change seems to be one of the top drivers, but there are also other factors that may increase the risk of microtoxins in our supply chain. And the way that we can identify these risk factors is either using the literature or end not not either using both the literature from all the studies that there are out there for predicting microtoxins, but also very important is to consult the experts and to hear from their experience to trust their expertise on which are the risk factors that we need to take into account. What is important to keep in mind when we are talking about the risk drivers and these factors here is that behind each factor in order to integrate this factor in our model, we need to have data. So again, the data, the importance of having homogenous data that can help us to build such models at a large scale is very critical. The other principle, and this is something that's well known in the world of AI is that it's important to compare different machine learning and deep learning methods, and to select and to choose the best performing one, the one that fits better to the specific problem. I mentioned this already. So, if we want to keep one thing here is that we should not go straight and directly select one method because we feel or we think that it will perform well, we need to compare different methods and choose the best performing ones. How to choose the best performing ones is can be answered through validation. So another principle, a very important principle is how is that we need to validate and we need to select a very good validation framework. The validation framework can be based on accuracy criteria, accuracy based criteria, but it's also very important to use non accuracy based criteria like for instance in the case of if we want to optimize the monitoring program. It's important to use a criteria like the associated costs of monitoring. So we can use also this criteria to identify which is the machine learning method or deep learning method that performs better than the other methods that we can use for this specific problem. So validation, it's very important to keep in mind that validation, we should use accuracy based metrics but it's also important if we want to have a practical application of the AI model to use also a non accuracy criteria. And here I give the example of associated costs. So let's see using these principles how a dashboard and an AI dashboard that can be used to set up a risk based monitoring approach can look like. So first of all, I hear also in the things that it's important to consider when we are developing such model is that they should be user friendly and they should provide allow some interactivity. So we, the goal here is to provide an interactive dashboard, allowing us to select specific mycotoxins, as Maria mentioned it's very important to select aflatoxin or carotoxin different types of aflatoxin or even emerging mycotoxins. And also to select the specific food or feed for which ingredient for which I would like to have the risk prediction. The model is able using also the regional information, the geographical information is able to highlight the most risk risk regions, and by selecting a region for instance United States, India, China, or Ukraine. We can go and see specifically for this region, which is the level of the risk. When it comes to the data. A very a first thing that we can do a first, the first thing that we can use in terms of data are the large data sets that we have from the testing from the monitoring programs and there are several variables that with high value there. The country of analysis the country of foraging the sample ID, the ingredient, the data, data of analysis, the specific mycotoxin but also the analytical result so it's important and we can use this kind of data. To allow us to predict if we will have a non compliant sample or subsequently bats. What such an AI model could predict them. The AI model in this case if we use, for instance, the monitoring results could predict the number of the batches that are non compliant and should be analyzed. These are the true positives and the false positives, and those that should not be analyzed, which are the true negatives and false negatives, and this can give us a very good idea about the probability of having a non compliant but for instance for maze for having a non compliance non compliant but for aflatoxin in maze from Brazil. So we can have already the risk level and the probability of having an increase risk of mycotoxins of aflatoxin in in the batches. As I mentioned it's also very important to transform this the prediction results to something that is practical and very useful for the companies and so here is the cost of monitoring program, and there is a recent work by one get all that we used a very interesting way of transforming the prediction results to the cost of the monitoring program to the dissipated cost of the monitoring program by using the follow up actions that we want to apply in each case. So this is, this is something that we can use in order to deliver through the dashboard, the estimation of the cost for the program for the monitoring program, and this can be done based on the predicted results. For instance, we are showing here that the for the true positive. It's important to do the follow up actions is of course sampling and analysis. And maybe storage. So, these are the cost components also that we can use to estimate which will be the cost for sampling analysis and storage for the one batch that it seems that will be non compliant based on the predicted results. And this kind of results can help us to, to see also the return of investment to see how good is the optimization so how, how, how much better, and which is the cost of the risk based program compared to the current monitoring program that we have. One last thing about the practical application of the model is that the, as already Joe mentioned is very important to provide an answer to the question, which is the sampling and analysis approach that I should follow. And of course they are best practices for that from in regulation. And these are following their own private system for monitoring microtoxins. We know that increasing the number of samples or collecting samples at multiple steps or at multiple control points can decrease the probability to accept contaminated but so taking into account all these measures, we can build a very good sampling and analysis strategy. We can use also this is an optimization problem there, how much we can test in order, which is the approach that we should follow in order to make sure that we will make the risk of microtoxins minimum. So an optimization model again can be used here to define the sampling and analysis strategy at the different stages of the supply chain. And that last but very important thing. And this is a good practice that we are following in our AI dashboards is the transparency and the explainability on how this model works, which are the data that has been used for building the model, which are the economical model that was used to estimate the cost, which are the most important insights that we see. So I will finish my presentation by saying that using adjusting the monitoring program and making it a risk based from being reactive to proactive, but of course, there are also other measures that can be activated, like proactively adjust the supplier practices, like to adjust your audit plan, and maybe in some cases also change suppliers for some period. So thank you. Thank you very much for your attention. I would be very happy to answer questions that we, I see that we already have in the Q&A, but also I would like to hear the reflection from our experts from our panelists here. So doing that, we can, we would really appreciate to get your answers about which are the food and feed categories that is important for you to have such a risk based monitoring approach. I will give a minute for that and then Chris, I will hand over back to you so we can continue the discussion and to hear the reflection from our experts. So thank you very much, Janice. We'll give, we'll give everybody some time to cast their vote. And as always, when I hear you speak, I learn a bit more, but also more questions come into my mind as well. There is already a lot of questions coming into the Q&A box. So please add some more questions. I think there are some really superb questions there that I'm going to pose, particularly to Janice, a little bit later on. So I'm really just head back just to see how is the voting going on and have we, have we found what the most important categories are, Janice? Yes, so I will close now the poll. I think that, and I will share also the results with all. So I think that cereals and nuts and seeds are the ones that it's very important to have such risk based monitoring approach. Of course, based on the, on the answers of the participants and thank you so much for that. But also other categories like animal feed, like herbs and spices, and fruits and vegetables are very critical in terms of categories. So thank you so much. And I would love to hear the reflection and then to answer some questions from the audience, if possible, all of course. Great, Janice, thanks. And I think that poll is really good, because I think it really accurately reflects a lot of the information we see, let's say, and rassive notifications that it's spread across many, many different categories, but cereals and nuts and seeds are really, really important. So I think we can now move on to reflections in terms of our panel. And really, again, we've had some time to listen to your presentation to think about the presentation. Again, we go back to you just in terms of, I guess, the three key questions that we posed you is, do you agree about what do you think is currently missing? And what value do you see in this AI approach for industry? Thank you, Chris. And Janice, thank you for a very thorough presentation and helping us understand what this AI based approach would look like and how we come together. This intelligence would really enable microtoxin prediction to be leapfrog to go to the next level. And this is about everything and anything pertaining to this issue as a whole system. So the type of microtoxin, the crop type, where they grow, where the food is consumed, whether it is direct application of the crop or indirect into the food, consumption patterns, trade patterns, and equally importantly for the company's investing, and actually for national labs and academia investing in this kind of research is how do we maximize the outcome for the resources invested? How do we get the best possible information? How can we create the best possible predictive models, which would then lead us to early action and effective action to mitigate aflatoxins? Equally, for those who do not have unlimited budgets, the question is, this is the budget I've got, what can I possibly do for the problem account? This is very important, particularly for small and medium sized enterprises. What is missing? Some of the things that are missing today, I don't think they will be missing in the future because technology and AI and digital and technological infrastructures are really growing very fast in industry these days. But I do think that it is important to have universally reliable and accurate something. Samples have to be representative and the analytical protocols have to be accurate and precise. What we're talking about here is we're talking about heterogeneous systems like grains and nuts and fruits. They are not homogeneous in nature. When we homogenize everything as much as you're done at the level of the crop. Also, as we get more and more data and we produce multiple iterations, the models will become better. I would like to see more incorporation additional to sample analytics. I would like to see the incorporation of other data sources, like for instance, the Internet of Things, traceability data, meteorological data, and so forth. But this will come also with the sophistication and the confidence that we will build in these models. The last thing is a repetition of what we said earlier on, we are dealing today with the catch work of risk tolerances, as well as regulatory limits. So this, if we had more consistency and more convergence, I think it would be easier. It would help with the prediction and both of the ability of the model, but also I think analyzing the cost, what it would take to build a model. Definitely there is a huge amount of value, the ability to predict and have better foresight and insights extremely important as early in the supply chain as possible. When we identify and address the problem and you're on, we're having multiple benefits, including less environmental impact because we're not shipping the grains, we're not shipping the product. I think it is very important to uphold quality standards and compliance to the regulations and actually the promise of brands to consumers and customers. And I think strategically at an industry level there are so many benefits. The ability to anticipate and address emerging patterns is very important because then you can have the right coalitions between the private enterprise and public sector and academia to actually address issues pertaining to the people, the planet, but also the economies involved here particularly when it comes to trade and very valuable income for people who grow the crops and particularly people growing them in developing countries and areas where climate is really a big threat right now. Maria, thank you very much. I have to quickly go on to Joe now and also give you some time to reflect on what we've heard Joe. Great. Thank you, Chris. Look similar to what Maria has said is probably a lot of synergies between the response here but I think certainly we all agree from an industry perspective, a robust predictive model to look for the risk factors and to predict future issues around how we collect the toxins is definitely needed. Collaboration between the regions and obviously the finer details around how you collect that the validation of the system. And I see some of the questions touch on this as well and I want to leave time for those, but I think the robustness of how we collected in a standard way how we sample in a standard way, how we get better engagement globally really in that it's a globally sourced material, and obviously the prevailing conditions where you source it from have impact, and obviously how we collected in a format in a standard way to inform on the risk profile that we can rely upon as a robust measure of risk from microtoxins. I suppose what's missing maybe the approach and the system validation and kind of was touched upon but I think obviously the one my colleagues on the on the line here as well from the data analytics side of it. It's just to review that in detail. We all are familiar with analytical methodology validation ISO standard validation of the same is just to again in AI speak. And I agree with Maria the more data we have the more robust and reliable the data will be in the prediction so again it's just to look into that a bit more detail. The value that I see for the industry is the pre warning that it will bring. It's a game changer really in relation to how we manage the business. And obviously, rather than relying on testing and you know information from in market of course that will continue as normal to maintain food and feed safety. But this will be the next level of prediction that again can be a true global model on microtoxin risk so I'd encourage that for sure. Joe many thanks that that's a really good good summary and I'm going to be really quick just in terms of my reflections because I think you know Maria Joe and I agree that there's there are lots of positives. I think also we're agreeing that we need to see more evidence about about case studies really that those predictive analytics have these desired consequences. In terms of the value for industry. I think Janice the data that you showed is you can make a real business case out of this based on financial returns for for investment in new technology. So I think I think there's lots of things that that we can we can as a group reflect on, but I'm really much more interested now not on my reflections but on the audience actually so maybe. If we can think about getting ready to take some some questions and I think there are really good questions here and I'm going to start off with. And I'm sorry Janice you're going to be in the firing line here the questions are coming your way. And they're really good questions and in terms of the source of the data and the models to feed the risk prediction. In terms of the data that you get how robust, can you say the different data sources are from academia from industry from NGOs. Have you any way that you can moderate the data that goes into the models. Yeah, the only way currently this is this is a very good point and I fully agree, Chris that does not make sense for us to make the reflection. That's why I'm going directly to the questions from audience. And so this is a very good point there. We are applying two ways of moderating cof ensuring the quality of the data. First of all, I want to answer on that that we need to have a review so we cannot trust the data that are publicly available, not that they are not good data they are good data but there are many problems with the different formats used with missing values in some cases so we need to moderate them. We are applying two ways. The first way is machine algorithms that are checking which are the gaps of the data, if there are any missing values, if there are any non consistent values, or if there are any values that are not according to data standards. For example of data standard this for instance, the class classification for hazards, or the food x classification for food products, and the other thing that we are doing is the after this machine curation and machine quality check we have also experts on our side that are checking if there are any issues and they are validating the quality of the data. And in my according to my experience this happens not only for the public data sources but even data that we have inside the organizations. We still have the issue of harmonizing called the data of having data gaps. So the same methods needs to be applied there as well. Thanks jealous. I'm going to go on to because often people used to think that it was 25% of all crops were contaminated with mycotoxins and I said no it's more like 60 or 80%. So the reason we've moved from 25% to 60 to 80% is about two things. It's about better sampling better testing, but also the really big factor is changing our climate. So obviously, how can you really start to think about incorporating climate into your models because climate is a big thing and and you know, it can have massive effects in in very small isolated region. So how do you start to think about building that in to really build the robustness of the models. Yes, so adding starting only from one very critical parameter like for instance the all the monitoring results that we have globally is a good start but of course the models needs to include also other factors. Adding a parameter on for the local weather is something that is very important to do so. And there are methods machine learning methods that allows that and we can add this parameter. So combining more traditional models like the empirical one, the empirical ones, or the mechanistic is another way of integrating more information about the cropping system about the soil type. The harvesting conditions. And this can be so we can have a hybrid approach that is not a hybrid approach in terms of mechanistic using mechanistic and empirical models but in terms of combining the first predictive models with machine learning approaches. In order to add and to take into account also other factors like the weather the climate change the geopolitical situation. And many more other many factors that we can, we cannot. I'd like to follow up on something particularly think Joe mentioned, and that was around validation. And you know I think we're all I hope pretty comfortable about validation of analytical tests in terms of how you determine their sensitivity their accuracy and so forth. We're actually talking about a very different type of validation Janice because you said is you train your model and then you validate your model, but to my mind, you also have to test that model to make sure that your validation has actually worked as well. And again this is a is back to the robustness does that that does that training and validation. Does that mean that that percentage reliability accuracy that you talked about is true, or is it just a kind of a mathematical hypothetical number that you're coming up with. This is very important what you're describing in terms of validation and also Maria and so mentioned that is that having only the mathematical validation so using the accuracy metrics and taking advantage of the, of the knowledge that we have of what happened in the past, and I think what the model could predict with what actually happened mathematically will give us the result of the accuracy and but this is, this is the one thing. We can do this in an exhaustive way so we can see different periods for different cases and we can see mathematically how good these models are and which is the performance of the models. But it's also important to validate the models for the specific application that we want to have. So even if we have a very accurate mathematically if we have an accurate model that does not perform very well when we want to have cost efficiency. It does not make sense to select this, this model so there is a trade off there of selecting the models that can perform well from a mathematical point of view using only the accuracy oriented criteria, but also to consider the non accuracy criteria, like in this case, the associated cost of the monitoring program and how much we can make it more efficient or fit it to the specific budget that the company has to make it very efficient for this amount of budget for this budget is very important. But it's, it's a very, it's a very critical step in developing the models I fully agree with your point. And I mean, I, first of all, I want to thank everybody for all of the questions that came in and there's many, many more questions that we don't have time to answer. And hopefully, I mean, I'll discuss with Janice will maybe try to get you some answers coming back after the webinar because I think the questions are absolutely fantastic. And also, I think, you know, one of my real learnings about AI and actually, I was working in AI 10 years ago but I didn't know it was AI, it was just, just good computer science is that just because you think you have access to one good machine technique, just be wary because it probably will not give you the best results or more most reliable results. And that's why I don't know has this real unbelievable toolbox of machine learning deep learning as well I think that's really a lot of your intellectual property as well. So I hope everybody enjoyed the webinar, I most certainly did. I really want to thank Maria Joe Janice for really wonderful content and discussions. I just want to thank you all for joining. I think it's been another really fantastic part of learning for me and I hope you gain something for as well. So it's goodbye for me and from our panel just I think a quick way of to everybody who joined us today. Thank you. Thank you, Chris. Thank you, Maria. Good job. Very much appreciate it. And you are all invited if you want to participate all the participants you are invited if you want to experience first time the microtoxin does but then provide some feedback, because we are still developing things in this, you are very much invited to express your interest. So thank you. Thank you so much all really appreciate your time. Thanks. Bye.