Jonas Christensen 2:42
Shalini Kurapati, welcome to Leaders of Analytics. It's so good to have you on the show.
Shalini Kurapati 2:50
Thank you so much, Jonas, for having me. I'm really excited to be here and I look forward to this podcast.
Jonas Christensen 2:55
Well, it is going to be an interesting episode and I know that because you are a person with a very interesting background. And you're going to tell us all about the art and science of MLOps and we'll get to that in a minute. But before we get to that, we want to hear a little bit about you. So perhaps, could you tell us a bit about yourself, your career background and what you do?
Shalini Kurapati 3:16
Absolutely. I'm Shalini Kurapati. I'm originally from India. I did my bachelor's in mechanical engineering in the College of Engineering Guindy in Chennai. And right after that, I moved to the Netherlands for my masters and I wanted to have a more broader perspective of engineering and not just going in technicalities of it. So, I did my Master's and PhD in technology, policy and management, which means that I was looking at how technology affects society and the wider business and the processes around it. And after that, I entered into this world of data, because in the year 2016 - 2018, big data was kind of exploding in terms of potential and I was quite interested in how data was being used. And my first, let's say, interest into data started from the whole management perspective of it. How do we manage all this data? How are we going to use this data? So, my career kind of started in the field of HR management, stewardship, and kind of the whole on the data aspects of things. And soon after, given all the data, AI was really quite a rational next step because what do we do with this data? We can extract meaning from it and we can do fantastic things with it. Powerful new models. And together with my partner, we started a company on AI. We decided to just make use of data. And we had really a consultancy company in the Netherlands, where we said, ''We're going to build models. Give us the data. We'll build your models''. It was quite interesting because this was in 2017, where we didn't even do much marketing and we got a lot of requests of ''Hey, build us a model''. So, that's how I started off with data and AI. I'm here with Clearbox. But just to give you a little bit of perspective of why I moved from doing building models to the current startup that I have, Clearbox, is that when we were building these models, everyone was just excited to use AI, because AI was a buzzword. ''I want an AI model''. We were building it, giving it to them. But we noticed that many of them were not actually being used in real business operations. It was just some kind of a project or an R&D or an innovation team, they wanted to experiment and many of them were staying as experiments. And we started looking into what was going wrong and in the whole process, we understood the multiple issues that relates to putting AI models in production and we wanted to do something about it. So that's why, with a few other very good people, we started this company called Clearbox. It helps companies to accelerate AI adoption and actually exploit the value of AI, rather than just experimenting it and saying, ''Hey, I use AI''.
Jonas Christensen 6:00
Very interesting. So, you've really done a full end to end data science experience, I could call it. Right from managing data to building user solutions to actually try and operationalize those. So that's fascinating. What really attracted you to get into this data science and AI? I know you've talked about data and and how you find it interesting, but what was the sort of thing that really made you decide to go down that path?
Shalini Kurapati 6:29
I would say that right when we were excited about the opportunities of data, we also saw the challenges that come with it. Like, especially in 2016 - 2018, that's when you also had a lot of scandals that came up, like the Cambridge Analytica about the misuse of data, about how AI is making decisions that people don't understand and how it could affect in ways that we cannot comprehend. And I was quite triggered by the Black Box problem of AI. They're very powerful machine learning models, make decisions that humans don't understand and we still go by it, because we say that, ''Oh the computer cannot make a mistake''. However, we know that computers or the AI algorithms learn from the data. The data is not perfect and it really reflects how we as humans make decisions. So, really the Black Box problem triggered me to say that, ''You know, we have this powerful technology. There are so many unknowns about it. There are so many issues about it. We should be able to extract value in a meaningful and responsible way''. And that's why we named ourselves Clearbox to get away from the Black Box problem.
Jonas Christensen 7:36
I think as a human race, as builders of AI around the world, we have a long way to go in dealing with all these ethical conundrums around AI and most of them only come about when the problem has already happened and the tip of the iceberg is big enough, so to speak. So, I can definitely relate to what you're seeing there. I want to hear more about Clearbox but before we do that, what I've noticed in doing my research on you is you've actually lived around the world pretty much, in so many countries. So you must be a bit of a globetrotter. What makes you jump from country to country, seek out new experiences?
Shalini Kurapati 8:13
I'm amused by the word globetrotter because it seems like a pre-pandemic term. But, no, I had the opportunity to live in a few countries. Mostly because of my Education and Professional requirements. So, I moved to the Netherlands for my master's studies and while I was doing that I had the opportunity to work in Volvo in Sweden for my master's thesis project. We were working on technology and trust. Like, how do companies build technology while trusting each other? And then during my PhD, I had the opportunity to spend a few months in the US and the University of Maryland. And then, I also spent some time in Germany and then most recently, I've moved to Italy. So, I would say I've lived a lot more in Europe in my adult life. And my intention to live in these places was largely driven by my professional interests and also my interest to explore how people do their business in different countries. And I had the opportunity and I took it. I'm really excited that I could do it.
Jonas Christensen 9:16
Europe is a good place for that. I didn't appreciate the proximity of things when I lived there and I grew up in Europe, of course. And I live in Australia, but the fact that I could go on a train for 45 minutes and be in a different country is not even close to being reality in Australia. There's nothing nearby here. It's a very big place.
Shalini Kurapati 9:34
But you have the amazing Outback.
Jonas Christensen 9:36
Yeah, we do. Now, Shalini, let's get to hearing a bit about Clearbox AI solutions. So, you are the CEO and co-founder of Clearbox. And can you tell us about the Clearbox platform and what problems you solve for users?
Shalini Kurapati 9:51
Yeah, absolutely. So, Clearbox AI, we are a young startup. We are about two years old. So our vision is to enable trust, for the AI in enterprise. We do it through our product. We call it the AI Control Room. So, what the AI Control Room essentially does is it helps data scientists or data science teams who build AI models to standardise their procedures of building AI models, make it more efficient and we create a series of operations that are easier for them to assess their data, to validate their data and to put the models into production in an efficient manner. So, what usually happens in the whole data and AI lifecycle is when you start building the model, 80% of the time is spent on data collection, cleaning, validation, and only 20% is spent on models. So, during this entire process, it's quite ironic that it's a lot of manual labour involved in making the process automated, so that is scientists, or even the data science teams, they spend insane amount of time in cleaning and validating if the data is good enough to put it into production. And usually these steps are not scalable from one project to another project. So if they work on data cleaning and putting into production from project, they have to redo it from somewhere else because the standards in this field are quite new. This is a new field of building AI models and putting into production. There's a lot of inefficiencies in the whole process. So, we really help them to reduce the time spent on putting the models in production. And to make sure that they also have a kind of model and data registry, so they can always go back and check which model perform better. But what really happens in AI cycle is there's a lot of trial and errors. So, when they build a model or when they look at the data set, they kind of tune some parameters by hand. They can't really keep track of which parameter worked better when. So, we also give them an opportunity to keep track of their experiments. So, it's really to decrease inefficiency and to have some kind of reproducibility of the requirements and make their life really easier. And the other aspect I would like to mention here is, I talked about building AI models as experiments. Nowadays anybody can do it. You can just buy a model of the shelf from the internet and train your - You put your data in and you will get some results, whatever that is. But the real issue lies in using that model and integrating it into your business processes. So, what do you need? You need a team of very highly skilled people to make that happen. You need specialised infrastructure. You need specialised tooling and you also need to have an IT knowledge of making this happen. So, usually you have data science teams, IT teams, then you will have, let's say, the business owner who needs to use this solution to improve their processes. And all of these people need to come together to actually put an AI model into production and there are many steps in this process that makes it very inefficient. And we kind of streamline those processes for at least the IT and the data science teams.
Jonas Christensen 12:58
That's a really good overview. You talked about how Clearbox enables robust, explainable and monetizable AI models. So, how does this actually work in practice?
Shalini Kurapati 13:08
So, our Clearbox product ''AI Control Room'', which if you look on the internet, you can even access it. It's actually based on our proprietary technology, based on generative models. So we have an AI model, it's called ''Variational AutoEncoder''. If you're interested, please look it up. It's a fascinating technology. So, we have a proprietary implementation of the Variation AutoEncoder. So, what we do is if we have a model and a data set, we create fictitious data points around modelling the dataset to enhance the datasets, to enhance insights about what we can learn from these datasets. And we also, using this synthetic or fictitious data, we can also create more explanations about the decisions. Through our engine, we provide not only data augmentations, we also provide explanations on why a decision has been made or not. Like we already talked about the Black Box problem, it's very hard for the person who has developed the model to even understand why a model took decision or not. So, we provide explanations. We also help to validate the data set. And if dataset is error-prone or weak or incomplete, our synthetic data can augment this data set and make the predictions more robust. Actually, there is a prediction that in future the majority of data science projects will be based on synthetic data because it's very difficult to get hold of it. And then we also have monitorable because we constantly check with our technology if the performance is stable or not. So, what happens is when you test a model in with your laboratory conditions where the data is static? With the synthetic data you can augment your data set. So in case your data big or error prone, you can augment it, improve the data quality. You can generate explanations from the model decisions, like getting explanations for the powerful machine learning models based on neural networks are very difficult. And then we also have the model registry. So especially for reproducibility purpose or audit trails, especially in the financial industry, you really need to know which data was used to make which model, in case there is a compliance request. And also, we also make it monitorable. So, what happens is in laboratory conditions, when you build a model, you get a certain accuracy. But when you put it into production, you have new data, then performance might not be very good. So, you need to have a very efficient monitoring system that says that your model is not performing with the new data. What's the data drift there? So, we create these information on top of your AI model. So that's why it's more robust and explainable and trustworthy.
Jonas Christensen 15:43
You've described to us the use of synthetic data. And I think it'd be really good for listeners to understand just how important it is for the future and why it is so important.
Shalini Kurapati 15:53
Absolutely. So synthetic data is really computer generated data or algorithm generated data that reflects real world data. It can be from different forms. You must have heard about GANs, which kind of create a person that doesn't exist directly. That's one type of synthetic data. But we're talking really about kind of cloning an original data set and that represents the statistical properties of the original data set. But it has a better quality because you can augment it. And synthetic data is going to be the future of AI development for a number of reasons. Because having access to good quality data is very difficult. For a lot of organisations, I had already mentioned that 80% of the time required to build AI models is spent on data collection, synthesis and cleaning and validation. A big chunk. And most often what happens is many companies, they don't have standardised ways of having different streams of data together to gain the most value from their AI models. It's really hard to manage data and especially if it's multimodal, if you have different streams of data from different departments, you not only have technical problems bringing them together, but you also have organisational silos that might not allow different departments have one consolidated data set. And you can't wait until you have all the data before you start innovating. So if you want to start innovating with AI, your best bet to start off is with synthetic data. If you can have access to a small sample of data, there are methods where you can use that to clone into augment your datasets and start training your models without having to wait to get the perfect datasets from organisations. Because I would be surprised if someone from a company would say, ''Oh, we have all our data figured out and I can immediately give you to start building an AI model''. That usually never happens. So, synthetic data is the best bet to start, you know, accelerating with AI.
Jonas Christensen 17:54
Thank you for that explanation. So, what you're really describing for us is operationalizing machine learning. That is also what the Clearbox platform is there to do. And the term that we use for that these days is MLOps, and it's a bit of a hot term. Could you explain to the listeners what MLOps is and why we need it to succeed with advanced data science solutions?
Shalini Kurapati 18:17
So, MLOps is short for Machine Learning Operations. To put it very simply, MLOps are a set of best practices that companies should take to implement AI in their business processes and get value from it. So MLOps is quite vast. It covers the entire lifecycle of AI in operation. So, it starts off, again, with data, a big chunk of it is actually data collection, analysis. And once the person who is developing the model decides that the data set is good enough and they build the model and they maybe build multiple models and try to understand which model works well and then put that into production and then that model needs to be monitored. And in this whole process, you need a data science team and you need an IT team and you need a business team to work together. Make sure that the performance is meeting the business requirements and the performance is up to the technical standards and the data that's coming in of good quality. It's really a collaborative effort to make MLOps a reality. It's really a hot term. And a lot of people mean a lot of things. For instance Explainable AI, Trustworthy AI, all are right now coming into the umbrella of MLOps in terms of implementation. It's really about best practices of putting AI models in production.
Jonas Christensen 19:39
So what kind of roles and responsibilities are needed in that process from various functions in an organisation?
Shalini Kurapati 19:46
Like I said, MLOps is really a niche. It's a new up and coming field and some companies have nailed it. There are no standards and how it is. We can kind of compare it with what DevOps used to be for software a few decades ago. Like 25 years ago when there were no processes of how to develop software and use it and have an update and that there's an entire software engineering domain for that DevOps. So, it's a good analogy to the MLOps. MLOps is a lot more complicated because there is board involved with models involved and data involved. All the three together. But the main roles I would say, would be data science divisions of companies, IT divisions have to work together and then problem owner because AI, the success of AI is not just, whether it's a technical success in a company, it's also a business success. So, the business unit that really has the problem and has the data and say that you could do something with that, it should also be involved in the whole process. Then in addition to this, you also need a specialised tooling and infrastructure. Tooling could be something like an MLOps platform that automates this whole processes or they have to build the whole thing from scratch. So, they have to have also an organisational view of who does what and are we buying an ML ops platform to automate these processes or are we going to be so huge that we're going to build everything from scratch or hire new people and make sure that we have a data, IT and monitoring divisions working together with a business unit? It's a new way of doing business, and some companies are doing it. But it's really the baby steps and many companies.
Jonas Christensen 21:27
Yeah, and we often see that there is this, it's called the proof of concept to production gap. Data science projects struggle to move from initial concept to an operationalized solution. Is MLOps the one and only answer to getting success with AI?
Shalini Kurapati 21:45
No, I would never believe in one and only solutions. I don't think that that will ever happen. But MLOps is definitely a solution for reducing the production POC to production gap. I mean, interestingly, I read a report two weeks ago, McKinsey states that only 17% of deep learning models go into production. It's 50% for the entire AI models, but only 17%. While MLOps is a solution, there is also another issue here, because the ones who, you know, the business units or the model owners, they should have the right expectations. They should not assume that this is going to solve all their problems. So one of the issues, in addition to the trust issues, the technical issues and the monitoring issues is the expectations issues. What do you expect this model to do? And sometimes they think that's the snake oil that's going to solve all their problems and that's also a reason why some models don't go to production. And another big big issue, why models don't go into production or they're stalled for weeks or months is lack of data or they're multiple issues why they don't go into production.
Jonas Christensen 22:53
You just talked about data. Well, I suppose you have been talking about data all along. But you talked about data missing in model building and model implementation. And I picked up on that because I know you're a strong believer in what's called data-centric AI, which is to some extent in opposition to model- centric AI. Could you explain these two approaches for listeners and tell us why data centricity is so important?
Shalini Kurapati 23:21
So, while I'm a big proponent of data-centric AI, it doesn't mean that I don't believe in improving models. So, data centric AI is an approach to building machine learning models where you keep the model constant and improve your data. While model-centric AI is an approach where you keep your data constant and improve your models. This term was recently quite popular because of Andrew Ng, founder of deeplearning.ai, who is proponent of data centric AI, but I'm a proponent of this also. Because what happens if you focus on data-centric AI is that you will end up with very good quality data, which means that even if your models are not as powerful, you can still extract a lot of information from it. Whereas if you just keep improving your models and forget about your bought your data, you might end up with very complex, very powerful models that are computationally intensive, but they might not even perform half as good if your data was actually fine. And this is also a matter of equity, because nowadays, you have these massive AI models coming along and everybody wants to use the most powerful, the biggest deep learning model, the most complex deep learning model. However, they have to kind of start with a baseline model. With good quality data, even your baseline, a simple AI model might outperform the most complex AI model. I completely believe in the fact that garbage in, garbage out. If you have bad data, you're not going to get excellent results, even if your models are perfect. But the problem is getting your data right is less fancier, less sexy, if you may compare it to building more cool models. So, that's why I'm always saying, ''Let's get your data right, even before you start thinking about building your fancy models''.
Jonas Christensen 25:11
Yeah, and back to your 80% of the work goes into preparing the data, etc, for the model. I think part of that percentage is actually getting the data right, because it is the most important part of getting the output. So, I couldn't agree more. So, we've got all these models and we're using MLOps, what are the best practices within that framework that businesses should use to get the most value out of their AI?
Shalini Kurapati 25:39
I think the best practices would be for them to start small, because they shouldn't view, especially companies if they want to integrate with AI, they shouldn't see AI as some kind of a vanity project that's going to give them good PR and buzzwords. Obviously, that's a good side effect, of course, but they have to start small. First of all, look at what they have. Look at their data infrastructure. Look at their IT teams, their capabilities, and say, ''Hey, let's start small and start building with simple models, with baseline models''. We also in our platform, we always give them a baseline model to compare with others. And see how good this is performing and then start building a machine learning model or a deep learning model and then check how good they are and constantly check with the business unit. Because you can build the best model and your business unit might have no value for it. And also check is there a need for a powerful model? Or what can we do with the data right now? Or what more data do we need to make it better? So, this kind of continuous monitoring of how the models perform are very important. They have to be iterated. They cannot be one big project. Let's say that, ''Hey, let's just build a massive AI model and try to solve all our problems''. I try to approach constant feedback from the technical and business units and really help them to accelerate and also create trust in the process.
Jonas Christensen 27:07
Nice. So, what are some of the most common AI pitfalls to avoid in this process?
Shalini Kurapati 27:13
Yeah, one of the common pitfalls would be to, it's also related to the previous question, it's obvious that AI is going to bring a lot of prosperity. And right from the beginning, companies should also view it with a pinch of salt. So, what we notice is that either people are super excited about AI and say, ''Let's just do it'' or they're super scared about it, saying that, ''Oh, my God, I don't know what to do with this'' or ''It's too complicated'' or ''It's a lot of unknowns''. So there's a kind of a polarising view. There should be a middle ground where they have to acknowledge the potential of AI and already try to understand what are the risks that come with it. The production gap is more business risk but also a regulatory risk or the privacy risk. They have to already anticipate and understand ''How do we deal with this, so that we'll have a strong foundation?''. So, they need to set a very strong foundation and again, starting with their data infrastructure: Get your data infrastructure right, make sure that your data pipelines are well oiled and once you have good quality data, you can build on it. And some companies, if they don't want to risk it, they could also start working in a simulated environment, especially in the healthcare and financial industries, where regulation is a big problem. They could work on Regulatory sandboxes. And if they don't have data to start with, start using synthetic data. So, there are many ways that companies - The biggest pitfall is to avoid vanity projects and build a strong foundation to build AI. Because what we notice is that the way some companies do it, they do it as a one-off thing, as like one project, but not with a long term view. Because AI is there to stay, so, they have to start building very strong foundations to scale AI.
Jonas Christensen 28:58
Yeah, I often say to people that they need to avoid the allure of using AI as an outcome in itself. It's a capability, not an outcome.
Shalini Kurapati 29:07
Indeed, indeed, absolutely. You've said it.
Jonas Christensen 29:10
Yeah and what you get out of that is firstly let's see if is the right solution. There might be an easier, more straightforward solution to the same problem. But also, if you want to do AI, there is, which is what you've raised, there's a lot of ancillary setup that you need to do, especially if you're in a regulated industry. So, let's pick on financial services, as you did. You need to consider the ethical implications. You need to consider what you're actually allowed to do with your data that it involves customers. How you can use it for models, but also what are you allowed to do with the outputs subsequently. How are you going to implement in your organisation, etc. So there's a real chasm between dabbling in AI for the fun of it versus doing it for actual business operations that is probably not appreciated yet to the extent that it needs to be. Hopefully, this podcast is one of these places where people can learn that. That is definitely my goal.
Shalini Kurapati 30:04
I hope so. It's really again about setting the foundations right. Like you said, that it's really a capability. And I would like to repeat that it's not just people skills, but you really need both specialised infrastructure, organisational change of mind and also the right expectations. What can I do for you today, in the near term and also in the long term? Because they should have like a vision. And maybe right now, you could have a solution that works quite straightforward. But in future, as you have more data, how are you going to streamline it to make good use of it? It starts with strong foundations.
Jonas Christensen 30:36
Yeah. And I think companies should really consider not just what they get out of it today, but what they can learn and how they can create this organisation learning curve, because it's a little bit like where AI is now in data science. More broadly, is akin to where we were with personal computing and in IT in the corporate world, in perhaps early 90s. And there were companies back then that said, ''Oh this is just a glorified typewriter. Why can't we just use our old typewriters?''. To some extent, that might have been true at that point. But fast forward 10 years, and all of a sudden, you're 10 years behind. Fast forward 15 years, and you're gone. So, there's a need to keep up with the trends, whether you like it or not.
Shalini Kurapati 31:21
No, no. I mean, it's true. But also, we also see a lot of disparity in the level of adoption of AI among companies now. For many companies, AI is chatbot. You know, ''Oh, we have a chatbot that's enabled''. Okay, good for you that's useful for customer care. But some companies, we, for instance, work with a financial organisation, they have a massive data science team and they work on MLOps principles and they have set up a very good process to discuss their models, and they have all these capabilities, but they are an exception, not norm. And some companies are really about ''What is AI adoption, what is production?''. You know, we see a lot of disparity. And I noticed, especially in Europe, compared to the US, Europe is definitely slightly behind in AI adoption in general. I don't know how it is in Australia. I'm curious.
Jonas Christensen 32:11
We are also behind the US, I would say. Probably also behind Europe.
Shalini Kurapati 32:18
Interesting.
Jonas Christensen 32:19
Yeah. It depends on where you are. I think there's no doubt that the US and China are definitely the world leaders in this space
Shalini Kurapati 32:26
Israel too.
Jonas Christensen 32:28
Yes, and Israel. Quite advanced and the number of people I meet from Israel that have AI startups and products in the data space, solutions in the full data pipeline is quite remarkable for such a small country. Now, Shalini, we touched on a little bit the human element of all this and how humans interact with the AI that we produce. In the last few years, we've heard a lot about human centred design. And this concept has also now reached AI development.
Shalini Kurapati 32:59
Absolutely.
Jonas Christensen 33:00
Can you tell us what human centred artificial intelligence is and why it's so important to have this?
Shalini Kurapati 33:06
Absolutely. In the context of human centred AI, is essentially for humans to have control over the AI systems. That's number one. So, there's this concept called human agency. So, you have this fully automated system that's making decisions for you and you as a human should be able to intervene and control when necessary, so that the system doesn't override your intervention. And it's also an ethical concept, because if you have a fully automated decision making, if something goes wrong, who is accountable and is there any remedy to decision that has been made? And this comes both from an ethical and also a design and as well as a legal perspective. So, of course, Europe is the world leader in regulation. So, in 2018, there was GDPR, the article 22 of GPR was mentioning that if you have automated decision making, if you use personal data for automated decision making, the data subject has the right to request explanations, request human intervention. So, when you build your systems, you have to have mechanisms or tools to let the humans intervene. It's not quite straightforward in AI systems trust and control are very big issues because you have this completely automated system, where you don't understand why the system is taking the decisions and how do we you employ these controls is a big research topic actually. And the other aspect of human centric decision making or what we call is ''The human in the loop'' machine learning, which means that there's also a new field of study called deep learning, where a human kind of interacts with the system and teaches the system when it goes wrong or has mechanisms to flag if the system makes an error and the human is able to go and teach the system how to do better next time. We consider this to be more human centric. And also, we firmly believe that we should design AI systems where the system should also be able to say, ''I don't know'', because it doesn't always need to give a decision because they are not infallible. They are as good as the data that they get and the way they have a lot of intrinsic and extrinsic biases based on the model or the person who's built it or the data that they get. So, they should also be able to say, ''I don't know. it can happen''. So, all these aspects we consider a part of human centric decision making. And also, human centred AI should also provide trust to the user. The decisions that they're having are fair or legal, for example. What's interesting is the European Commissioner actually released a set of principles, upon the principles of trustworthy AI, which people who build models of businesses should adopt to make sure that it's human centred, it's of course, trustworthy.
Jonas Christensen 35:56
Yes, so there is a big element of risk management in this, of course, and also making sure that we consider ethics and fairness. Those two terms, by the way, in many cases yet to be defined. That's perhaps the 2.0 episode of this podcast conversation. Now, a little bit of a left field question, we often hear about AI becoming very, very strong and powerful and being able to do things that we're not capable of even dreaming of as human beings. Do you see that can happen without this regulatory framework for human centred artificial intelligence or is that just hype and Hollywood movie material?
Shalini Kurapati 36:40
As of today, what you're talking about is the artificial general intelligence that might or might not take over the world. I would say as of today, it's hype. We are not even close. We are not even close as of today, that AI will take over. But 50 years from now, maybe. Because we have some of the best brains in the world working on making it possible. For instance, most recently, the founder of DeepMind, has started a new company for drug discovery and they are focusing on improving the capabilities of artificial general intelligence, where is this all performing AI. For the moment, we only have Artificial Narrow Intelligence and we are not even able to get that right at the moment. So I'm not a doomsday heir or a naysayer. But as of today, it's really a hype. Few decades from now might start to become a reality. But we are really far off.
Jonas Christensen 37:33
Okay, so let's enjoy the next 50 years then, before it happens.
Shalini Kurapati 37:38
Regulation is definitely important because we need to respect fundamental rights as humans and also our existence as a human race. But we also need to look at it from like a risk point of view. Right? What are the consequences of these new technologies? It really is a big subject. And we should also not be super scared of it, because we can extract a lot of value from these technologies. But it's also a matter of mindset and how do we extract value from these systems as powerful as they are. Like, we can do different things with nuclear energy. You can create power, or you can destroy. It's how we as humans are going to exploit this technology. Regulation is definitely one thing but there's also a lot of geopolitical issues happening because Europe, for instance, is high on regulation. For instance, other countries might not be as much. So, what are the ethical implications if one goes ahead and the others are behind and what are the powerplays? So, there's a lot of issues in addition to ethical aspects.
Jonas Christensen 38:38
Absolutely. There's a lot of politics in this and I think your analogy of nuclear technologies is really good. So, if we consider the most advanced form of AI that we haven't invented yet to be the potential nuclear holocaust and what we have today is still these regular accidents of nuclear power plants blowing up or we spill uranium here and there and poison a few people. Of course, I'm referring to AI here doing things we don't want it to do and we only find out once the accident has occurred. And we still want to be able to regulate that and make sure we have the right protections in place. Now, Shalini, just taking a broader perspective here, as an expert in developing and deploying AI solutions, what are the some of the world's biggest challenges you think we'll solve with AI and was some of the most exciting developments on the horizon?
Shalini Kurapati 39:33
There are many but from our personal experience and also our recent work also with the United Nations Working Group on healthcare AI, I firmly believe that healthcare will be one of the fields that will benefit a lot from artificial intelligence. So far, we hear a lot about the use of AI in healthcare, but there are very few examples that have been put into practice and used continuously beyond radiology and a few other. In terms of drug discovery, in terms of pandemic predictions, in terms of access to healthcare in remote areas, I think AI can really improve the access to health care, affordable health care to a lot of people. We are just scratching the surface of it and I firmly believe that that's going to be fascinating.
Jonas Christensen 40:19
So what are some of the use cases you see there?
Shalini Kurapati 40:22
So, for instance, one use case that we've noticed is, like, in terms of virtual doctors who can help nurses in remote areas to check ecographies of pregnant women where they're not able to reach them. So, we see that. And there's also using the data science prediction of the risk of pandemics. So,how they spread. For instance, the forecast of the next pandemic. We also see diagnosis. So, what happens - Like, we are working with a hospital in Belgium, where they have a lot of data from different departments and they would like to already predict a patient's chance of getting a certain disease, even before they get it so they can give them the opportunity to advance their tests, to improve their patient outcomes. There are so many use cases. It could be as simple as an AI app that can help people to diagnose, let's say, skin cancer at their homes, if it's done properly. The use cases are endless.
Jonas Christensen 41:19
Fascinating and somebody's early detection where you use AI to take some of the attention away, as opposed of making it easy for people to perform the task is really fascinating.
Shalini Kurapati 41:30
Yeah,
Jonas Christensen 41:31
It means you can only detect. I have no doubt that we can develop something that's better than Dr. Google.
Shalini Kurapati 41:36
Yes.
Jonas Christensen 41:38
Personally, I've been diagnosed with an incurable deadly disease at least 10 times in my life from Google. But...
Shalini Kurapati 41:43
Oh, yes.
Jonas Christensen 41:44
That's a story for another time.
Shalini Kurapati 41:46
Indeed. No, no, absolutely. The healthcare AI is really fascinating also because the risk factor is also very high when you use AI. The fairness and ethical issues are also very high. But with great power comes great responsibility. But we do have some very good use cases and we are actually working if you're also interested, you can look up it's called aiaudit.org. It's an international group. It's coordinated by the United Nations and WHO, with startups like us, academic institutions. But what we are really doing is we work on a few healthcare use cases and we kind of develop frameworks. How do we assess this AI model for robustness, for expandability? It's in the early phases but we see a lot of potential in the near and far future.
Jonas Christensen 42:31
Aiaudit.org.
Shalini Kurapati 42:32
Yeah.
Jonas Christensen 42:33
I will check that out. Now, Shalini, we're coming towards the end here and I only have two questions left for you. But before we get to dose, is there anything in this space that you would like to get across that we haven't mentioned yet?
Shalini Kurapati 42:48
Although there are many, I've had fun speaking to you about, even today, we covered a broad range of topics. So, I don't have anything particular than what we said.
Jonas Christensen 42:58
All good. There is no need to fill the space if you don't have that. So, my last two questions. So, the first question is, I always ask the guests on the show to pay it forward, because we are a global community of data and analytics enthusiasts here. So, who would you like to see as the next guest on Leaders of Analytics and why?
Shalini Kurapati 43:20
I have many names, but I actually thought about it and I think I would like to see Ivana Bartoletti. She is the global Privacy Officer of Wipro and Wipro is the big Indian software company and she recently joined as the Global Privacy Officer. So, what essentially does is try to understand the implications of regulation and privacy in building AI models. So, I would like to know her experience of actually implementing that in our business environment. So, I would like her to be here.
Jonas Christensen 43:20
Yeah, fascinating. That is definitely a conversation I would like to have too and Wipro, as a company is also really fascinating. Huge beast, you could say. It's like a small city, almost. They have 300,000 employees or something like that. So, there will be a lot of solutions coming out of that place with various privacy challenges to consider. Now, last question, Shalini. Where can people find out more about you and get a hold of your content?
Shalini Kurapati 44:27
Well, about Clearbox itself, our website is the most updated place we have. It's clearbox.ai. We have blogs. We have regular interviews with folks. We update our product features and we have a lot of interesting material on social media on LinkedIn or Twitter. These are the three main platforms where we are active and we also have resources, white papers and links of information about us and there is also a free work flow tool. Anybody can sign up and just have a look.
Jonas Christensen 45:02
Yeah and I did just that. So, I would encourage listeners to go and have a look on there. So, I had a look at your LinkedIn page. There's some really interesting articles on there, especially around synthetic data, which I learned a bit about and within the platform as well, I had a trial run at your product. And I could see the ease at which you could click around and do things. I haven't uploaded my own models and done anything as far as that but I really do encourage listeners to go and check it out. It's very easy to use and you'll quickly get a sense of what you can do in there and I think it actually helps to, perhaps demystify some of the elements of MLOps to some extent and make it less daunting, because you've got the technical setup for it there.
Shalini Kurapati 45:47
Absolutely. Thank you.
Jonas Christensen 45:49
Shalini Kurapati, thank you so much for being on Leaders of Analytics. I really appreciate the time you've taken to teach us all a little bit about MLOps and the past, present and future of AI more generally. All the best for the future for you and for Clearbox and I can't wait to see what it all turns out to be and hopefully we'll still be alive and friends with AI in 50 years.
Shalini Kurapati 46:14
Thank you so much, Jonas. It was an absolute pleasure speaking with you. It was a very friendly conversation and I very much enjoyed it. Thank you very much.
