Machine Learning & Artificial Intelligence

Virginia Robbins (00:04): Hey guys, welcome to PCBBB’s new podcast, Banking Out Loud. I'm your host, Virginia Robbins, and here with me today is Sonia Port Worth, my co-host. We're both members of PCBBs management team. Hey Sonia, how you doing? Sonia Portwood (00:18): I'm doing just fine, Virginia. Thank you very much. Excited to be here for our first episode. Virginia Robbins (00:23): Yeah, me too. Yeah, now. Hey, Sonia, it's really clear from your accent where you're from or where my folks may think you're from, but mine, well, it may be a bit hotter. Well, a small hint there, . But if you'd like to guess where we're from, there's an email on this page to send you in your guesses. And I hear that, Sonia, that marketing's got some cool prizes, should somebody guess, right. So while you're doing that, look for our teaser. I've been sitting on the edge of our seat waiting for this first episode to drop, and I think it's kind of fun to go and check it out if you haven't yet. So why don't we dive right in, Sonia, we have decided to pick this little small topic, one that's easy to understand, and the rest is our first one. It's machine learning and artificial intelligence. (01:05): Well, yeah. Okay. I am being sarcastic here, but it's by no means a small topic. It's constantly changing, it's evolving. But before we jump in and introduce our guest speaker today, Hey Sonia, did you know that artificial intelligence and machine learning is actually older than ATM machines? ATM machines date from the 1960s? And the science of machine learning and artificial intelligence were, they came from the 1950s and besides giving Arnold Schwartzenegger some great catch raises and answer questions and mitigate fraud today, they also help us make better cookies and quite frankly, do a whole lot of range of other things. Sonia Portwood (01:44): Aware of that, nor has the other things you just said. But that's amazing because I know ATMs, I've been in quite a few of those and they've been around for longer than I have. So yeah, that's a long job. Virginia Robbins (01:55): So the, the thing with AI and machine learning is we see these terms thrown out by vendors. We see fintechs claiming that the, these AI and machine learning increase efficiency and provide better customer services. But how do they really work? What's going on here? And I think when you listen to these, this episode and the one that follows, whether it's a customer, a new employee, maybe a board member who stops and talks to you about where your institution is using these technologies, is this something we need to jump on? Today's topic should help you answer those questions, should make you look good and give you ideas to look at maybe the current challenges you're facing a bit differently, maybe to decrease expenses, increase revenues, maybe improve the quality of your team's work, or help you just have a different perspective Now. I know. Sonia Portwood (02:40): Yeah, it'll go ahead. I was just gonna say, it may actually bring up more questions for you, which will your interest, and you can look into those. I know that as we were talking to our guest ahead of time, it certainly raised some additional questions for me, so I like that. Yeah. Virginia Robbins (02:55): Super. So we, Sonia, we've talked about our guest today, so it's time to actually introduce him. We've got with us Shiva Sandy, as our guest today. Shiva, thanks for joining us. Shiva Sandy (03:06): Thank you for having me. It's a pleasure to be here. Hello, Virginia Robbins (03:09): Shiva. Hello, Shiva. You have a master's in artificial intelligence from MIT, is that right? Shiva Sandy (03:15): Technically my master is in human computer interaction. My bachelor's is in ai, yes. Virginia Robbins (03:22): Okay. So my bow heads down, bow down to the king of geek here. But what I love about Yeshiva is that when we talk about this stuff, I know that you're also focused in on where the value is and how to apply these things and bring them real. So really thank you for being here today. So before we get started, let's do, uh, 20 seconds. Uh, let's learn about Shiva. So Shiva, I'm gonna throw some questions out at you. Just shout out the answer and we'll go on. So, Shiva coffee or tea? Shiva Sandy (03:49): Tea Virginia Robbins (03:50): Beach or mountains? Shiva Sandy (03:51): Beach. Virginia Robbins (03:52): Summer or winter? Shiva Sandy (03:54): Summer. Virginia Robbins (03:55): Take out or cook home. Shiva Sandy (03:57): Cook home. Definitely. Virginia Robbins (03:58): Pizza. Pizza or burgers? Shiva Sandy (04:01): Pizza. Virginia Robbins (04:02): First job in banking. Shiva Sandy (04:04): Technically working for TD Bank when they were me with Commerce Bank, leading that exercise. Virginia Robbins (04:10): Wow. Cool. Favorite class at MIT Shiva Sandy (04:13): Statistics than probability. Virginia Robbins (04:17): Better you than me. and favorite thing about machine learning and ai, artificial intelligence. Shiva Sandy (04:24): My favorite memory about it was while at MITI had a professor tell me that while it's cool to look at neural networks using neural networks, you'll never be able to drive a car, have the computer generate text or prose that reads like a human or to do visual identification. So I learned from that. Never say never, Virginia Robbins (04:52): Never say never. Yeah. There Shiva Sandy (04:52): The reason why he said that and he believed it. And what happened in the time sense to make all of these things possible. Virginia Robbins (04:59): So as you said, this is a hugely evolving topic. So time to jump in. Let's start talking about what is machine learning and artificial intelligence. And, and Shiva maybe you could help us with define what these things mean from, from someone who studied them. Shiva Sandy (05:13): So I think of artificial intelligence as a technology or using a technology where we can create intelligence systems that can simulate human intelligence. We talk about intelligence. We really want a machine that mimics cognitive functions that we think of as humans. So learning and problem solving. Machine learning is a subset or subfield of artificial intelligence, which enables machines to learn from past data and experiences without being explicitly programmed. So machine learning algorithms or systems must be trained on data. The more data you provide the algorithm, the better it gets. And then there's a subtopic, deep learning, Google and Amazon uses a lot with a subset of machine learning where you use neural networks to process vast amounts of data. And any neural network that has more than three layers is considered deep. So I think the goal of AI is to make a smart computer system that solve problems like humans. Machine learning is to teach machines with data, how to do a particular task and get an accurate result. And with that comes lots of successes, but lots of things you need to look out for. I read somewhere that 90% of the data right now was created in the last five years. Wow. It may not be true, but I I think we're somewhere close to that idea. So we, one of the reasons it's useful is we are using more and more data and it can be leveraged effectively. Virginia Robbins (06:53): So there's a, there's an aside comment here about my email box. 'cause I feel like I've 90% of that data sitting in my email box. But ish, when we talk about data, what is data? Is it interactions? Is it transactions, is it all of the pieces? When you think about data and this, how should we think about data? What is data? When we, in the context of Shiva Sandy (07:11): Machine learning, I would say all of the above. Okay. It it depends on how you want. So again, right, the difference between AI and machine learning. Machine learning, and people try to divide AI into what's called weak AI or strong ai, machine learning is solving a type of a problem. Okay? Right. So if you want to understand speech or natural language, you have data related to those elements to solve that problem. If you want to look at lots of transactions to identify either the patterns for fraud or predictive analytics, your data subset is there. So it, it depends on the problem you're solving. That's how I would think about what's data and what's being applied to. Virginia Robbins (07:56): So anything that we capture digitally could be the data that we bring into the problems that it's dependent on the problem we're trying to solve or just the, it's any organized data that we bring into the problem stack. Shiva Sandy (08:07): So I think anything we capture could be data. Right? Right. Um, machine learning tends to work best on structured or semi-structured data. AI on the whole can use unstructured data elements. But you said with digital, I guess even if you capture paper, you can easily convert that paper to images and then feed that into the system. So I would say anything captured can be used as data. You may need some pre-processing. Virginia Robbins (08:36): So, Shiva, why as a banker do I care about this? This sounds all interesting, but it sounds stuff that the really big companies are doing. Why do I care? Shiva Sandy (08:44): So, so we've all lived through the last year and the pandemic and the constant lockdowns have boosted demand for digital services. And the future really needs to be powered by AI to keep up with the growing demand for cutting edge technologies in financial service. That means customers come to expect faster service, better service. Right. It studied by McKenzie says that year 2030 AI expected to save or add $1 trillion in savings or added revenue. And that's only less than a decade to go. Currently 32% of banks currently use ai and as you alluding to, banks have always had to innovate and keep up with technology. Mm-Hmm. . Yep. In the 1960s we had ATMs in the seventies we had card based payments. Mm-Hmm. in 2000. Online banking then mobile banking AI is basically the next step to work with. And to deliver on the needs of today's society, you have to integrate with ai. Ai ha offers numerous benefits to finance. You get automation and automation can help reduce costs and increase revenue. Sonia Portwood (10:02): Can you give us examples of where this is already showing up and to our customers, whether it be through other banks or other fintechs or anyone out there, just some examples of FinTech that our audience would already be familiar with. Shiva Sandy (10:18): Right. So examples for right now, when I think of ai, um, slash machine learning in financial services, banking automation as mentioned is a big topic. JP Morgan uses RPA robotic process automation and AI to do a lot of their regulatory work and to create reports in there. Crest Financial, a major leasing company uses AI and machine learning for credit decisioning and some type of analytics work. Bloomberg has recently come out with an AI powered tool to make trading better. One of the larger hedge firms out there in the US uses AI powered technology for sentiment and news analysis. So reading social media as well as news reports to try to understand how certain traits can be made. Risk management is another area that Crest Financial does to, um, figure out as they make their financial decisions how to reduce risks. Plaid, which is used by a number of the large banks and financial institutions use AI and machine learning to do fraud protection. (11:26): Most of the major credit card providers also use some type of machine learning or AI to identify rogue events and to reduce fraud. And the one that I think everybody is most familiar with is a vision of a personalized banking. So Bank of America, I think they use Erica. And again, these are visions of if you have a smartphone, Siri or Alexa or hey Google, a chat bot or technology that allows you to interact with a system. Um, for those of you who taught, Erica was a real person. I'm sorry, it's not, it's an AI . Sonia Portwood (12:03): Yeah, that's helpful. It gets us all sort of on the same page and make sure we know exactly what we're talking about here before we move forward and into a deeper dive. Thank you Sheva. Virginia Robbins (12:12): Well, Sheva, you talked mostly about the large banks, but we see smaller institutions using chat boxes and going through that. So Shiva Sandy (12:20): I try to identify, um, from, well-known names, but yeah, the nice thing about AI and today is lots of these technologies are what I call commoditized. You can get them off the shelf or partnering with fintechs to implement at your institution. If we want to consider just the after the possible with what a commercial bank or communication could do right now, you can get intelligence, personalization and some type of omnichannel. So in preparation for this, I I try to sketch out what I thought some of our customers could use AI to do. Mm-Hmm. . Just imagine you are a small business owner and the bank, you, you the business with a bank and the has integrated, um, and understands your company's cash flow and how you're doing. They can send you a customized lending solution or loan option in there just because they've analyzed your data and use AI to find the right option for you. (13:21): Then you can with a few clicks and using your face for recognition, for telling the truth, you can be approved for funds once you get the funds. Um, systems can also help you as a customer from the bank, manage your inventory or work through some of these. The bank can also hook you up with the right partners for your business to be successful. If you're doing lots of the business with the bank. Some of these services could be preparing your taxes on a quarterly basis and definitely giving you an AI advisor. So you interact with your bank. You either call or use their them different channels and they understand your business, they understand your relationship and they can give you personalized advice at that point in time. Virginia Robbins (14:07): So, Frank, for instance, managing loan covenants, which can be a very tedious activity today. You could look at your data from that and build a machine learning example around that, for example. Yeah, Sonia Portwood (14:18): I think that those are great examples and I think the future is unlimited as to where this could show up, especially as we gather more data and organize our data better. I think that the uses are, are limitless going forward. So why don't we go ahead and jump into the different types of AI out there and try and put those in some categories that may make sense to our audience and then we can see what may fall into which bucket. Virginia Robbins (14:47): Sounds great. Shiva, you wanna take us through different categories? Shiva Sandy (14:51): Right. And I'll caveat this with, this comes with my bias and . This may have, and that's something we need to look out for AI bias, but I tend to segment it into natural language processing, computer vision data and analytics and robotic process automation as the four main buckets. When I think about technologies that are somewhat commoditized or easy to partner with folks out there that small community financial institutions can start using today to enhance their productivity, reduce their cost, and add value to their members customers. Virginia Robbins (15:33): Okay. So Shiva, you touched on a little bit about bias and that's something that in artificial intelligence has certainly been in the news. Why don't we take a step back before we hit into bias and, and talk about some of the challenges with artificial intelligence in addition to bias. Are, are there other challenges? Shiva Sandy (15:50): Yeah, so again, there are challenges that come with implementing a new technology into the banking system. And some of these get exacerbated when you try to put in an AI system or partner with some folks. So some of these are your core and legacy systems and having data siloed or not quite appropriate, um, throughout the organization. That's why at PCBBB, we try to build the data platform to help our customers get the data, their data in one place, but having the right skill sets and finding the right partner also meet some of the other, um, problems with it. But with the machine learning system and an AI system, as we said earlier, the more and better data you have, the better the results you're going to get when training an AI system. But I wouldn't say you need to be discouraged by that. You probably need to reframe the problem, right? (16:50): And try, instead of trying to solve everything right away, the goal is I try to follow the 80 20 rule. Can I solve 80% of the problem with the data I have? We understanding that the system isn't gonna be perfect, but if I can reduce 80% of the volume or handle 80% of these accurately, I can hand the other 20% off to folks who I can either clean my data or the system can learn as I'm working through. So, and again, the advantage of these systems are they will learn over time to go through that. There are a few other challenges, uh, I would be remiss if I didn't mention some of the algorithms are not what I would call traditionally validated, right? So as you put data in regulators like to know if you're using a system, how do you get your results? Mm-Hmm. (17:38): Coming out the other end, right? And depending on how you are using it or what you're trying to use it, you will have to, um, work with the regulatory body to understand what their, their tolerance is for these things. Because some systems don't give you the same transparency and traceability and that needs to be fully understood. Um, the skillset for maintaining and supporting any new system is something that needs to be planned and dealt with. And technologies are evolving and changing. So you're going to have to keep up with it and understand that if you start with something today, make a business case so that within about three years you have an ROI 'cause you, you may have something better or more efficient in three years time. Mm-Hmm. That, so those were just some of the ideas. I would also think there might be a, a cultural element that how do you introduce the idea to your organization to get, to move from a, a traditional process to, okay, we want to be an AI set leading bank or a data-driven institution. Virginia Robbins (18:47): I can see some bankers just being concerned that, hey, I, when the, when something, when the cash question wasn't answered right, I could go over and talk to Sue or John or, or whomever, but if the machine doesn't answer it correctly, how do I know that and how do I manage that? So Shiva a bit, this also gets us into the discussion of bias. So how do you, as technologists explaining this to non-technology people, how do you help people understand to trust the machine? , it sounds like something out of, out of, uh, Terminator, but uh, how do you trust that and how do you deal with bias? How do you watch out for that? Shiva Sandy (19:20): So with that one, I go back to an old computer science too. Garbage and garbage options. Garbage. Yeah. Virginia Robbins (19:27): Yep. Shiva Sandy (19:27): Right? But it's easy to say that it is not an easy problem to solve. Mm-Hmm. , you need to look at the data and try to identify where there might be biases in that data and what sources you're getting from that data. The Brookins Institute actually did a study, and I refer people to go look that up on the web of, in ai, lots of the places where you can get bias in there. And technologists have been working on it. When speed recognition systems first came out, if you weren't white and you weren't male, most systems didn't understand you, , because the guys developing it were white males. So the data they gave in to educate the systems were white male voices. And then I think one of their bosses or somebody else was a female and she tried to use the system and it didn't work and they had . (20:20): So we need to be a little bit more cognizant of that going out front. But IBM and a number of different, um, firms have come up with frameworks and governance ideas to help you identify the data and identify biases in your data. And again, machine learning, it's a little bit easier to see, okay, what's the model and how the machine has come up with its model and what it, some of the elements it uses when it comes up with the model. If you use a neural network for deep learning, that gets to be a little bit harder. But to that end, you would go back to some of the governance and frameworks that people have identified to see what are we doing and how are we using to train the systems? And I, I may be a little bit conservative, but I, I tend to advise folks, let's crawl before we walk, walk before we run, start small, um, try to get value as we, uh, work through and identify some of these problems. We're gonna see where our data is good, where our data needs to be augmented and how we can get better as you implement this Sonia Portwood (21:24): Solutions. I believe that I've heard you use that crawl before you walk, walk before you run, not just in this scenario, it seems to be your motto and most of your IT ventures. Shiva Sandy (21:33): Yeah. We're engineer thought we can always do better. We can always do that. Let's, let's add value first. And Sonia Portwood (21:41): No, I, I think absolutely great advice and almost, almost all new ventures. Virginia Robbins (21:47): So Shiva, is there, is there a part of AI or machine learning, uh, technology that's a little bit more mature, a little bit easier to do? If I was listening to this podcast and I wanted to try something out, is there someplace that I could start that would be easier? Shiva Sandy (22:02): Uh, so I think, um, and again, I'll readily admit my bias and ask people to push back and test on it, but we've been using Siri and Alexa and Google natural language processing. And if you had a call center or some something where people called in to interact with staff at your institution, you have policies and procedures, um, that guide and can easily train a system and then it mm-Hmm. It's leveraging a couple different knowledge bases. The knowledge base in technology of fintechs to understand language. Mm-hmm. and to do it. But also in your system, if you have highly codified rules, it's easier for the technology to pick up these rules and use these rules to automate. So then you can go from offering nine to five service to 24 7 service. And that helps you build to a personalized banking system. I, I think today you, you shared, um, there are a couple credit unions that have implemented chat bots and digital assistance. (23:04): So the, the nice thing about technologies as it matures, it democratizes the barriers to entry for using these technologies get lower and people can put it into their system. So think about what do I do or what we do in our daily life that is somewhat rote or routine Mm-Hmm. that can easily be picked up. Right? And then if, if you marry that, the 80 20 rule of 80% of the time we're doing these steps, and it's not an exception, you can easily introduce automation. And when I think automation, I think, okay, can I get machine learning or any type of AI with, to add additional value in there? So the, Virginia Robbins (23:48): So for instance, uh, sorry, uh, go ahead, Sheda. Shiva Sandy (23:51): I was going back to your emails. , I would think that there are AI systems or that I can open emails, analyze the text of an email, file it, or if it's important, like, um, let's say it's for doing applications that come in, they can scan the critical part of the email, put those in, and even try to do some type of adjudication based on the information there to process some of these. Virginia Robbins (24:22): Well, I was just thinking that for a lot of banks covering after hour support, we can support ING customer calls in that period is, is a little bit of a challenge. And there may be a way to play with, since there is already probably a recorded message or someone else who's covering that, to throw in some additional tools that might help customers after hours or after support be a way of improving service with not much cost and a way to experiment. So I think there's lots of different ways to think about things that we might already do that might be places where we've compromised on either on delivering service, on the customer journey, on ways that we're interacting, we might, we wanna rethink about again. So I, I think this is really interesting that we, it's not only good to look at where the problems are, but look at things we might have compromised on in the past because either the costs or the technologies just weren't there. Shiva, we talked a little bit about bias and we talked a little bit about the challenges. Is there, uh, anything more on bias and the concerns on bias that you'd wanna share or any of the other challenges with ai? Shiva Sandy (25:25): Yeah, so again, I think we've, we may have heard stories that when Facebook or Microsoft tried to do an AI system, it quickly went from nice and polite to, to, Virginia Robbins (25:41): Uh, yeah, they took it off, I think. Yeah, it was like 24 hours later it was cursing at people. Um, Shiva Sandy (25:46): Yeah, yeah. And, and sping, um, other views. But that's one extreme, right? The other points is, okay, if we've made decisions in the past and that's the only data that we have, are we self-selecting out of all the possible customers that we could be using to put items in place? I I would also take it back to the having the right data. I've been stuck at home for the last year with my kids and they get out to school before I do. So they get onto Netflix and they've totally confused recommendations of what you'd like to watch. And again, it's, it's a bias. They've overwhelmed the system with their choices. So then when I get onto the system, it's recommending, um, cartoons that, well, I have some slight enjoyment of it. It wasn't necessarily what it was recommending before. So again, how you put items in there can affect it. (26:43): Um, to the other extreme, we also need to be mindful from a regulatory standpoint. There are some questions we're not supposed to ask folks. Mm-Hmm. , when we're making some decisions, the machines can also infer the answers to some of these questions, right? By other data points. Mm-Hmm. , right? So if someone buys ary napkins and we're not supposed to ask them their sex, there can be a correlation between the agenda, right? Theory. So getting, if we're getting data based on your purchases or certain elements, we have to be careful of how we training a system and putting it in place. If I've only made loans in the past 25 years to sit in zip code or a sit in arm area and I'm expanding my footprint of the business, the computer doesn't know that and it's trained on from this data pool or this customer set. These are good customers. It doesn't necessarily, it's not a smart AI system as yet where it can know it needs a, a different data set. Eventually it can be trained with new data, but it has to overcome that. Sonia Portwood (27:53): But, so Shava, do you have, uh, another example you'd like to share with us about the problem that arises when you cannot ask certain questions? Shiva Sandy (28:02): I saw your reaction to the first one. So I think I'm going to, um, not, um, try to extemporatively come up with some of these answers, but I, I think . Virginia Robbins (28:13): So, so why don't we, why don't we try one that is, uh, more closer to banking? It has a lot to do with where people live and redlining is illegal. But if you start to tie in economic conditions and look at where, where a person's neighbors are or the value of non-related assets, you can very quickly, if you're not careful, fall into some unintentional redlining. And so one has to be very careful about working with a vendor who's bringing in external data perhaps and looking at that external data to make sure that it doesn't include any discriminatory characteristics. Um, sometimes yeah. Sonia Portwood (28:52): I think the key around this is the discrimination Mm-Hmm. discriminatory data that you may not have, right? Because you are not allowed to ask that for one reason or another. And that could affect the data and the outcomes. Shiva Sandy (29:04): Right? And I think IBM and the Brookins Institute, uh, are two examples that you can quickly research if, if you want to go through putting it in place. They, they've done good work on data governance and AI governance. Regulators also starting to wake up to this and put it in that if you are putting in AI systems, you have to be mindful of these ideas. So it is a topic in and of itself for full with a proper expert to go through their governance. Um, but avoiding some of these items. But again, if you're putting in any new system, you have to work risk reward and make sure you are doing it appropriately. But there are lots of advantages that remind that lots of the big banks have put in AI systems for a reason. And that reason is it drives savings but also increases revenue. You can get to personalized banking. I give you personally what you want when you want it. And the only way we can really do that is leveraging the power technology and AI systems. Sonia Portwood (30:10): Shiva, what are the easiest types of machine learning for our banks to implement? Shiva Sandy (30:17): And again, I, I would think items with natural language processing fairly easy. Maybe computer vision related systems tend to be easy, depending on how you define, you can tie computer vision with some type of, um, robotic process automation. When I think about what's easy to implement, it's what technologies are tech companies using that a couple years old? So Siri's been with us for, is it four or five years? Yeah. Like Alexa is in there. Yeah. So it, it not only has a large knowledge base outside of banking, so that technology for understanding natural language already exists. So all you're really doing is teaching it. Now that I get this phrase typed or spoken to me, how do I interact with it on my system? Sonia Portwood (31:09): And as far as what makes it one of the easiest types is that, is that because it's mature and it's more out of the box, more of a plug and play? Is that what makes it easy? Shiva Sandy (31:21): Yeah, it, it's, it's mature. Um, people have overcome a lot of the bias or some of the governance. Um, type is used in training it. So you're getting a system with a certain level of intelligence outta the box and then you are training it to deal with your interactions, right? Robotic process automation or computer vision. If I need to scan a document or an image and be able to process, or even if I want to do facial recognition and go leveraging that, I'll, I'll caveat facial recognition comes with some, they haven't solved all their biases as yet, but they've done a little bit more. But if you narrow the scope a bit with some of the computer vision type items, you can clearly use computer vision with robotic process automation to handle workflow items. So if you have paper or documents coming in, how do I read those documents and put them into other systems or if I wanted to write regulatory reports or other items, those are also driven by what is the data, how do I train a system to do it? And then how do I write prose or text writing prose or text is something technology companies have cracked the hard part of the problem on. So, mm-hmm, that's what I would think would make it easier. Sonia Portwood (32:41): So if there's one thing that you wanna make clear to our audience when they walk away from this podcast and something you think that they should know, what would you say to sum that up? Shiva Sandy (32:53): I would say AI is well within reach of all of our listeners. All of them are interacting it with it today in their everyday life. Putting it into your bank or your credit union is not a difficult activity. But again, I would never think about it in a technology term. I would think about it in how can I add value or how can I reduce cost to my organization? And when I'm thinking about it that way, it becomes another option, another tool in the toolbox to help me achieve that. It's, it's a tool that two or three years ago wasn't available to all of our listeners today. It is in a couple years time it will be ubiquitous. We'll be interacting with it and it won't be there. But for us to take advantage and to grow customers and to, um, grow our business, we need to start thinking about how can we use these tools? How can we start putting them in place? And don't be scared. Start small, get some successes, you'll get some failures or some learnings as I think about it and you grow and you get to the next step. Get to the next point. Sonia Portwood (34:06): Perfect. Thank you Shiva. Virginia Robbins (34:09): Awesome. Shiva, thank you so much Sonia. Shiva, great episode. Just wonderful. We hope everyone listening has really enjoyed it. Catch up on all our past and future episodes@bankingoutloud.com and don't forget to subscribe to be notified each time new content is added. Thanks so much for listening. Thank you everyone. Thank you guys. Bye. Sonia Portwood (34:29): Bye.