In this episode of the Connector Podcast, host Koen Vanderhoydonk interviews Amit Gandhi and Monish from Praxis Digital, a consultancy specializing in data science, AI, and machine learning for banking and fintech. The conversation covers Praxis Digital’s flagship projects, including an invoice‑payment prediction model that helps fintechs forecast cash flow, and a satellite‑and‑climate‑data platform used by an African bank and the UN to assess land‑ownership risk. The guests discuss data challenges, emphasizing the need for versioning and the emerging role of synthetic data generators and data‑augmentation techniques to overcome scarcity while noting potential bias concerns. Ethical AI is addressed through a focus on encryption, privacy, and avoiding high‑regulation sectors such as healthcare. The dialogue moves to operational best practices: the importance of MLOps pipelines, model registries, and monitoring to mitigate model drift. Emerging trends highlighted are a shift toward data‑engineering, model‑agnostic workflows, and deeper understanding of deep neural networks. Finally, the impact of ChatGPT on financial services is explored, from personalized loan offers and insurance products to hiring and customer engagement, with examples of question‑bank generation. Contact information for Amit Gandhi and the head of ML, Chris Darcy, is provided at the end.
Chapters
Welcome to the Connector Podcast, an ongoing conversation connecting fintechs, banks, and regulators worldwide. Join CEO and founder Koen Vanderhoydonk as you learn more about the latest available trends and solutions in the markets.
Welcome to another Connector Podcast episode, and today I've got 2 guests with me. Amit and Monish, both working for Praxis Digital. Guys, can you explain us a little bit where you're from, what you do?
Okay, sure. My name's Amit Gandhi. I'm the founder of Praxis Digital. So we're a digital transformation consultancy with a specialization in data science, AI, and machine learning.
What about you, Monish?
I was previously in the AI and machine learning space for quite a long time. So yeah. We are just starting off with some new stuff in the AI and machine learning space now.
Great, great. And I mean, new stuff always sounds exciting, but are you able to share some examples of machine learning and AI that you're working on and that potentially have been also transformational for the business?
Yes, so we've been in this space for about 5 years. So my colleague Chris, he heads up our machine learning AI space in terms of building machine learning tools and algorithms to solve business problems. Business problems, mainly for the banking and fintech side. So yeah, Manish has come on board recently to to yeah to to offer a new service around sort of machine learning ops. So in terms of the sort of case studies that we've been involved with, one of them was like a use case would be to take hundreds of thousands of invoices for a fintech and then put I think build an ML tool to predict when those invoices are going to be paid. So that really helps with. their sort of budgeting and forecasting. Yeah, for the client. Other use cases have taken huge amounts of satellite image data for African bank and also a— yeah, also for the UN in terms where we've taken huge amounts of satellite data, also added climate data and other different datasets and then made— and then fed that into a front-end system where they can mitigate their risk, they can understand the risk and understand what's going on in the land that they own. So yeah, that's just a couple of examples.
It strikes me, and obviously no surprise, if you talk machine learning and AI, then you almost immediately talk about data-driven operations. When you talk about data, you need to get resources. So how does that work? How can you help the customers to built the resources, or how does that work in practice?
There are 2 ways to do the data, you know, data resources and other things. Most of the machine— most of the machine learning applications require huge amount of data, especially for the deep learning scenarios. But modern-day solutions is not about more or less about data, because it's sometimes easier to find the data, especially if we are dealing with large organization, more or less. But the problem is to maintain the data, versioning it. So you have raw data, now you need to version it properly, and you make some changes on the small piece of data, and then you have to keep track of all the data over a long period of time. That's the really challenging part. Another way now, because of generative AI, we also have the idea to provide solutions in terms of dataset generators. From taking a small chunk of data, we can create a huge sum of large replicas of the similar type of data. Although it's fake data, but it's good to test your model. It's good to have a real-life simulation from a large variety of data. So the more variance in the data is available, the more variety and more capabilities we can create from different types of models and test it on the real market with the real data. So we are not necessarily dependent on data nowadays, thanks to generative AI. Yeah.
So interesting. So let me sort of repeat because I'm not a specialist, but what I'm hearing you say is that You can create synthetic data that can actually be used to teach the machine learning algorithms.
I'll just give you an example of it. It could be something like a data augmentation tool for image processing, for that matter. You can't capture all the handwritings of every person in the world.
True, true.
Let's say you have experienced some of the handwritings. What you can do is you can have a small skewed syntax for that handwriting, maybe 50% of noise you will— not 50, I'm sorry, like maybe 2% of noise you will introduce. You will tilt the image a little bit, you will tilt it right, right to left. You can create multiple characters from that. You know, you can thinner the line, thicker the line, something like that. This is data augmentation, and this can be done in a much more rapid fashion, uh, thanks to the recent progress in the AI space. But you can create multiple data like that. Yeah.
Have you not the risk that you sort of— when, whenever, when we talk about data and we talk about AI models, then sometimes we talk about bias? So what's the risk by doing this, that you create more bias in the models that you create?
Yes, of course we do. But the best part is the mathematics comes into the picture. So it depends on what type of data you are generating, whether it's a normal distribution, whether you want to generate a certain type of skewed data. You always take an inference from the real-life data. And it is where the mathematics will give us the leverage of— Finding the right dataset and curate the right dataset for the training of model.
Yeah, because talking about AI, you very often talk about AI for good, the more ethical and responsible part of AI. So how do you look at that? What's your take on this?
It's my personal opinion. I mean, I don't know whether Amit will agree or not, or you will agree or not, but—
Interesting, interesting.
My personal experience, is data encryption has to be the number one priority that we need to see when we get— we are getting the data. So for example, if another, another very interesting part I feel is you should not enter into the territories of the industries where data privacy protections are really high. For example, healthcare industries and something like that.
Mm-hmm.
But let's say assume that you went there also, the highest probability to, to achieve the most ethical practices in those data is to have a proper encryption. So you will not have the similar bands available, but your model is trained on, on an encrypted data rather than the real-world data, and only the end user know how to use that model. So, so this is, this is the perfect scenario where the users can train their own models and have their own personalized models according to their own data, and they have full control over it. That's, that's my take on ethical AI at the moment, at least.
Okay, so I mean, these are potential risks. Are there any vulnerabilities that we need to take into account when we talk about AI?
So for example, if you train the model, if you train in a conventional fashion, there is a huge amount of risk without machine learning operations and other things. If you don't have a proper MLOps pipeline and other things set up, it's a huge problem, especially for the large enterprises. It's not suggested at all. But let's assume that you have a good model. and you have a good machine learning pipeline, like new data keeps on coming in and the model keeps on training itself, and then you have proper metrics to check your results all the time, the biggest risk you will enter here is model drifting, which means your model will go outdated after a certain period of time. To mitigate that risk, the number 1 scenario we will say is having a solution of model registries, which are all the modern-day machine learning operation tool provides. You have a registry of all the available models, And you can see their accuracies, how many accuracies, how model is working on maybe not exactly real time, but yeah, partially real time way. And finally, you will choose whether you need to retrain your model or replace your model. But you have the whole registries of model, especially when you are using any pre-trained model like Llama 2 or something like that from the ChatGPT kind of a scenario. This will come a really handy scenario. So yeah.
What would you see would be the most exciting developments and trends in the world of AI? And I hope you're not going to say ChatGPT.
Of course not. None of us. Actually, the most important trend that I'm seeing right now is people are moving more towards data engineering part than the machine learning and other stuff. For example, data engineering precisely means to me is we are more more and more becoming more model agnostic, and we are becoming more and more, uh, generalized kind of scenario that are happening. Not generative AI or generalized AI, not general intelligence. I'm not talking about that. I'm precisely talking about, uh, it's, it's more like the scenario is becoming, uh, we, we are not over-reliant on the particular type of data, but we are more reliant on understanding the models. Even understanding deep learning model is a field of research that are really gearing itself up. So I really want to know if somebody is working on that field and coming up with the solutions. When they can understand the deep neural network, it's insane. We can have very interesting scenarios. So this is what is happening.
Well, on the other hand, I don't want to necessarily avoid the question about ChatGPT because I think there's a lot of things happening. But on a more general level, and we almost come to the end of this podcast, What do you see as the effect of ChatGPT on current financial institutions?
There are, there are a lot of things happening in this space. For example, now people are having personalized loans based on their financial histories. Let's say if I have access to your mobile SMS from the banks, from your cell phone, and if you allow me through your app, I know your financial status and then I can offer you a loan when you need it. You know, this is, this, this data can be fed to ChatGPT and ask the question right away. with how much loan is ideal for it. You know, ChatGPT has special plugins people are developing on top of it. Another way is insurance is also getting personalized. Customer engagement is completely changing, completely.
Why is that?
You don't need a— for example, from the education sector, if you ask me, like one of the projects that I'm working on personally was creating a lot of question banks for any exams. Once you feed ChatGPT the type of questions, it comes up with a new set of questions, which is I don't know whether it's copyrighted or not, but yeah, it's kind of cool. You're not reliant on teachers now anymore for creating a question bank for similar type of questions, or you can combine multiple questions. ChatGPT can do it for you. You just need to feed it right type of data and train the model, something like that. So there's a huge impact on many industries, including particularly financial institutions, their hiring process, their customer engagement processes, their, their core business operations, all 3 of them are definitely going to have some kind of impact by ChatGPT in coming days for sure, or already has come up, we might not know it.
Brilliant. And I would suggest that if people want to know more about that, they should contact you, right? Or one of you both. So where can people find you?
So you can either contact myself on amit.gandhi@practice-digital.co.uk or my colleague who's head of ML, and AI, and his name's Chris Darcy. So it's chris.darcy@praxis-digital.co.uk. Um, and yeah, he'll be more than— we'll be more than happy to speak to you in more detail. Um, he has a PhD from the Alan Turing Institute, and we have a strong team that, that graduated from there or already in— already studying, uh, the line of academics.
All right, thank you very much. I thank you both for joining the podcast. Thank you so much. And also thank you to the audience for tuning in for some more innovation in our industry. Stay tuned. Thank you very much.
Thank you. Thank you.
Thanks for listening to another episode of the Connector Podcast. To connect and keep up to date with all the latest, head over to www.jointheconnector.com or hit subscribe via your podcast app. Podcast streaming platform.
