S8 Favorite: The Language of Money: Re-Engineering Messy Transaction Data for the Era of AI Agents with Naré Vardanyan, Co-Founder & CEO of Ntropy

0:01 [SPEAKER_00]: It's always very hard, like you said, and especially I think when you're in the weeds doing it, it's harder, deciding how good your first sort of version should be versus releasing it, so that was the one thing that I was struggled with quite a bit because on the one hand, we're saying that we make the data better, which means our output should be good. [SPEAKER_00]: On the other hand, the more people test the MVP. [SPEAKER_00]: The more data we get, the more chances we have to make it better.

0:29 [SPEAKER_00]: Financial Services industry in general has seen billions and billions of transactions. [SPEAKER_00]: And here we are a tiny company saying, we can do better with a small sample. [SPEAKER_00]: My name's Narev Ardenian, I'm the CEO and co-founder and entrepreneur. [SPEAKER_03]: This is Code Story. [SPEAKER_03]: The podcast brings you interviews with tech visionaries. [SPEAKER_03]: Six, six months of moonlight. [SPEAKER_02]: It goes to the last and out of the back house. [SPEAKER_03]: Who share what it takes to change an industry?

0:58 [SPEAKER_01]: I know exactly. [SPEAKER_03]: She took many girls to get rocked. [SPEAKER_01]: Who built the teams that have their back company is its team's help each other. [SPEAKER_01]: She was proud of her team. [SPEAKER_03]: Keeping scalability top of mind. [SPEAKER_02]: that infrastructure was not even fighting it as a group. [SPEAKER_03]: Total waste of time. [SPEAKER_03]: The stories you don't read in the headlines. [SPEAKER_03]: It's not an easy thing to achieve. [SPEAKER_03]: To get yourself accustomed, try to begin.

1:22 [SPEAKER_03]: To ride the ups and downs of the start-up line. [SPEAKER_02]: To really miss it. [SPEAKER_02]: Not just about technology. [SPEAKER_03]: All this and more on Code Story. [SPEAKER_03]: I'm your host, Noah Labpart, and today, how Nari Verdangian built the great equalizer in financial data through multi-geo, multi-lingual transaction data.

1:50 [SPEAKER_03]: Nari Vardignan has a statistically unlikely story. [SPEAKER_03]: She grew up in Armenia without electricity during a war. [SPEAKER_03]: She recalls that her parents gamified the experience which allowed her to experience it much differently than the hardship it was. [SPEAKER_03]: Eventually, she went to work for the United Nations and her words so she could save the world. [SPEAKER_03]: Though eventually, she was disillusioned by how slow things move. [SPEAKER_03]: It was at that point where she shifted over to tech.

2:15 [SPEAKER_03]: Outside of tech, she used to love reading, but now that she has a child, she sticks to audiobooks, and she thoroughly enjoys art. [SPEAKER_03]: Specifically, 20th-century Russian Jewish artists. [SPEAKER_03]: When Nari started to travel abroad, she noticed that, for some folks, the ability to obtain things in life, like a visa or passport, was a given. [SPEAKER_03]: Yet, others were not enabled to obtain these types of things. [SPEAKER_03]: She set out to create the great equalizer through enriched financial data.

2:44 [SPEAKER_03]: This is the creation story of entropy.

2:49 [SPEAKER_00]: What entropy is, we're essentially building the data infrastructure for financial services. [SPEAKER_00]: And it means that any product that you're building where you need or require financial data. [SPEAKER_00]: By financial data, I mean, information about what a company or a person is spending their money on. [SPEAKER_00]: For all the products that you can think of, that require that information, you need to use entropy as non-dubined glare. [SPEAKER_00]: And here's why the information, the record of where your money spend, every time you have a card, buy something online, is kept within different silos, within banks and other financial institutions.

3:29 [SPEAKER_00]: Sometimes, you yourself can look at your back statement and not recognize who the merchant is, what does it say and not remember that you went and bought something. [SPEAKER_00]: And because of this, it's very hard to work with this data, especially at scale. [SPEAKER_00]: If we think about how is this data used today, every single product that really touches our lives for mortgages to carlons, to renting apartments, of credit scores, payments. [SPEAKER_00]: remittances every single crucial financial product requires this information in the state and we're creating a unified layer that will standardize clean and reach categorize this information for the world to use internally we call it a money graph as well.

4:13 [SPEAKER_00]: I grew up in a place I told you where a lot of things weren't like anywhere else and there was a lot of hardship and then when I traveled and started living in grew up in a broad I'd realized that you know for some people things are given like getting a visa you just don't need them if you have the right passport for most countries or being able to rent a department. [SPEAKER_00]: get a SIM card for your phone, get a loan to start a business, but then for others it's way more complicated.

4:43 [SPEAKER_00]: They're basically queuing outside of room where everybody else is already poor. [SPEAKER_00]: I like to bring that analogy. [SPEAKER_00]: entropy was born out of this idea that financial data is one of the biggest equalizers. [SPEAKER_00]: And if we want to change things from the ground up, we have to make sure that that building block is for energy use, especially because there's so many builders in the world today, and building product is being democratized. [SPEAKER_00]: There will be way more in the future.

5:07 [SPEAKER_00]: I think there's less than one percent of the world population that can coat today, and that's massively grow. [SPEAKER_00]: With the advances of AI, no code, everything else that's out there, that's so fascinating. [SPEAKER_03]: Let's dive into the MVP. [SPEAKER_03]: Tell me about that first product you built. [SPEAKER_03]: How long did it take you to build, and what sort of tools did you use to bring it to life? [SPEAKER_00]: We were building a product that had a massive challenge.

5:31 [SPEAKER_00]: We needed to train models that would read and understand financial transactions, like a human wood, like us at the clean, categorized, enrich. [SPEAKER_00]: but we've never really seen financial transactions beyond our own at scale. [SPEAKER_00]: Neither had our models. [SPEAKER_00]: So we had a big challenge for the MVP. [SPEAKER_00]: Where do we get the data? [SPEAKER_00]: Financial data is not publicly available. [SPEAKER_00]: It's very sensitive. [SPEAKER_00]: Most of the stuff that you can buy.

5:58 [SPEAKER_00]: It's very expensive and anonymized and not useful. [SPEAKER_00]: So we actually decided to create a consumer product essentially a website where people with link their bank histories and we would provide them value whenever they wouldn't recognize a transaction and they had a question about it, whether it was fraudulent or a wrong

6:24 [SPEAKER_00]: Well, we did that. [SPEAKER_00]: People obviously consent that that information is being shared and then you can use that to provide the products and services. [SPEAKER_00]: We asked them specifically if we could use this to do a better job at understanding what they're doing with their money. [SPEAKER_00]: And that's how we go to our first data set. [SPEAKER_00]: I'm trained at MVP. [SPEAKER_00]: P, which was a model that would take in financial data from one of the core US aggregators lab and clean categorize and return information about who the merchant is in the transaction and what the transaction is about.

6:54 [SPEAKER_00]: And that's how we got started. [SPEAKER_00]: There was 20,000 transactions that we had access to with user consent and had open banking, cloud, finicity, all this sort of bank connections not existed before who would never be able to do what we did and build an MVP. [SPEAKER_00]: So there's a lot of work and infrastructure that's been built before to make this possible. [SPEAKER_03]: Okay, so let's stay on the MVP for a minute. [SPEAKER_03]: With any MVP, you've got to make certain decisions and trade-offs about how you build that MVP.

7:21 [SPEAKER_03]: You talked about some of them at a high level, but dive into one or two of them of some of the bigger decisions and trade-offs you had to make around technical debt or approach or feature cut, and how you cooked with those decisions. [SPEAKER_00]: It's always very hard like you said and especially I think when you're in the weeds doing it it's harder, deciding how good your first sort of version should be versus releasing it so that was the one thing that always struggled with quite a bit because on the one hand we're saying that we make the data better which means our output should be good.

7:56 [SPEAKER_00]: On the other hand, the more people test the MVP, the more data we get, the more chances we have to make it better. [SPEAKER_00]: Our competitors, or financial services industry in general, has sent billions of billions of transactions. [SPEAKER_00]: Here we are tiny company saying we can do better with a small sample because we have the better technology and we have a huge approach so on. [SPEAKER_00]: I think it took a lot of back and forth between myself and my co-founder and CTO, saying, hey, we should release something that is maybe slightly better than what they're seeing as a row output on their bank statements, there will be a lot of mismatches and a lot of things will return incorrectly, but people will use it and I will take it from there.

8:41 [SPEAKER_00]: and we'll see how we build the next versions. [SPEAKER_00]: So that was the first sort of trade-off. [SPEAKER_00]: Do we really, something that's not great, but then it gets used and becomes better because the data and the effect of that, and we decided, yes, it was painful because we were in a lucky position and we had a lot of demand, and there were so many cool names that were testing the product and we knew that the results are going to be what we want them to be. [SPEAKER_00]: It was giving us an opportunity to build towards the version that would work.

9:11 [SPEAKER_00]: Again, there's this magic version, you return every possible piece of information about this transaction or about this account holder versus you return the most critical piece of information. [SPEAKER_00]: And you do that quickly. [SPEAKER_00]: And again, we went for do it quickly return the merchant information and the labels start very simple. [SPEAKER_00]: See how useful that is and then try to build up on that. [SPEAKER_00]: Afterwards, when people started using it, there were so many feature requests of what they wanted to be returned, but the format should be what connections we should work with.

9:43 [SPEAKER_00]: Financial services today is a lot of Lego pieces built together, so if you're building a data layer on top, you should be very selective and thoughtful about what you're integrated with, what do you want to cover, what's important, what's not. [SPEAKER_00]: The second phase of it was we were just drowning in different feature requests, different use cases, directions we could go. [SPEAKER_00]: It sounds like a good problem to have, but actually not that much because everything is an opportunity class and lack of focus is tough.

10:13 [SPEAKER_00]: You want to do everything you want to be horizontal, you want to serve every customer, but it's impossible.

10:22 [SPEAKER_03]: So you've got your MVP, it's working and you've got it to a point where you're seeing some progress. [SPEAKER_03]: How did you progress the product from there and mature it? [SPEAKER_03]: And I think to wrap that in a box a little bit what I'm looking for is, how did you build your roadmap and how do you go about deciding, okay? [SPEAKER_03]: This is the next most important thing to build for entry. [SPEAKER_00]: We had an approach here that was. [SPEAKER_00]: quite orthogonal to what most people were doing in the space and we thought this is why this problem has been solved.

10:52 [SPEAKER_00]: Because all that we wanted to serve certain customers and have focus, we also knew that we want to build the most general sort of financial data, understanding possible. [SPEAKER_00]: And if you think about financial transaction, you need close to a common

11:12 [SPEAKER_00]: You need to read it, you need to understand the description, and then you assign a label, saying this is groceries, and this is somebody's salary, and this is rent, and this is income, and so on. [SPEAKER_00]: It's very hard to do general. [SPEAKER_00]: You can do it for one specific field, or area, one specific domain, like, for example, accounting, and chart accounts, or a very narrow demographic of people. [SPEAKER_00]: This is transactions for creators on YouTube. [SPEAKER_00]: You can do that very well with very high accuracy, but it's very hard to do across the board and we wanted to build something that was any source of data, any location, any geography.

11:49 [SPEAKER_00]: That was our end goal because we thought that's what's most valuable and horizontally can power. [SPEAKER_00]: Today's product of future product. [SPEAKER_00]: To be able to do that, we were constantly balancing building for our customers today trying to be focused on the features they're requesting, first this also, for example, starting early on to take transactions from Latin America and work with some partners there so we could build multilingual models. [SPEAKER_00]: and that was a big bet.

12:16 [SPEAKER_00]: Everybody's focused on one geography because that's what makes sense. [SPEAKER_00]: We didn't do that when global very quickly and the product actually even our U.S. product became better because we trained models that took part in different languages, saw diverse data and got better at the long-tailed even in the United States. [SPEAKER_00]: And since then, we've made a lot of decisions that are not very logical when you think about them. [SPEAKER_00]: For instance, if you can mitigate mistakes by building a lot of rules, then you should do that because then your customers will sign and pay you money, it doesn't matter that those rules work.

12:49 [SPEAKER_00]: Only for various specific cases and kill your generality. [SPEAKER_00]: Again, we didn't do that, which means our product initially had a ton of false positives and we had to mitigate it without customers and [SPEAKER_00]: continue to between the models to make sure the intelligence was improving versus we were applying some patches to local problems. [SPEAKER_03]: So okay let's switch to team them. [SPEAKER_03]: So how did you go about building your team and and I'm curious what did you look for in those people to indicate that they were the winning horses to join you?

13:19 [SPEAKER_00]: It is the hardest thing to get right. [SPEAKER_00]: Everything, people think fundraising is right, having the right idea, market, timing. [SPEAKER_00]: I think having the right people is the toughest thing. [SPEAKER_00]: We had a very high standard technically because what we were doing wasn't really... [SPEAKER_00]: trivial and a lot of ways hadn't been done before and there was a lot of intellectual curiosity that needed to be involved in a lot of skill. [SPEAKER_00]: We definitely had a very high bar there but also million eye both knew that we need people who are very smart work very hard but don't take themselves seriously.

13:53 [SPEAKER_00]: and sometimes this can be taken to an extreme but that cultural element really made it possible to bring a group of people together that just love hanging out and building things and not stepping at nights and times because there's a lot to do. [SPEAKER_00]: We also started through the pandemic so a lot of the people that we hired couldn't eat. [SPEAKER_00]: Our first employee was in France, and Tarihi is still with us. [SPEAKER_00]: He's a brilliant machine learning engineer from NVIDIA.

14:20 [SPEAKER_00]: He decided to take a leap and join a two-person startup. [SPEAKER_00]: And we had to go through the hurdles of hiring something France, which I keep telling him we wouldn't do it for anybody else, but him. [SPEAKER_00]: hiring in France. [SPEAKER_00]: I mean some regulations and tax legislation is just crazy which is such a big problem because there's amazing talent everywhere. [SPEAKER_00]: The whole team is distributed. [SPEAKER_00]: We have engineers in Europe with the US and we do quarterly off-sides where we get together and work on product to whiteboarding.

14:51 [SPEAKER_00]: We had to do it because the pandemic and right now we're trying to set up local offices so people could need because everyone wants that and I think we're graduating

15:02 [SPEAKER_00]: as a team.

15:07 [SPEAKER_03]: So as you step out on the balcony and you look across all that you've built, what are you most proud of? [SPEAKER_00]: I'm never happy about how things are, so when I think about things I think about things that don't work, I mean it's very hard to focus on things that do which I should do more. [SPEAKER_00]: It's definitely the people we have at the company and the team. [SPEAKER_00]: I guess that it sounds cliche but very intelligent people who could do anything else in the world have so many opportunities and they choose to spend their time on a dream and a vision that we've put together and they do it relentlessly with a lot of focus, dedication, very, very proud of that.

15:47 [SPEAKER_00]: The core group as well, like I said, we've made many mistakes and we've had people. [SPEAKER_00]: who we let go, or people who laugh themselves, but the core team is something that keeps the company together growing and also what makes our product different. [SPEAKER_00]: Because of the hours that the approach, the creativity, the skill, also the humility, I think it takes a lot of humility to do things that don't work and continue doing them to a point that they start working.

16:15 [SPEAKER_03]: Let's flip the script a little bit. [SPEAKER_03]: Tell me about a mistake you made and how you and your team responded to it. [SPEAKER_00]: One of the biggest mistakes that I still make is not reacting to a wrong fit from hiring perspective fast enough. [SPEAKER_00]: That's cost us a lot of time and effort and emotion. [SPEAKER_00]: It's brutal. [SPEAKER_00]: I know there's a lot of layoffs on the market. [SPEAKER_00]: And I really feel for people who are let go. [SPEAKER_00]: It's a tough situation.

16:44 [SPEAKER_00]: But what I'm talking about is different. [SPEAKER_00]: It's when you hire people for a job. [SPEAKER_00]: You think they're afraid very quickly you feel they're not. [SPEAKER_00]: But then you continue dragging the conversation because it's a hard conversation to have and you want to make everything work. [SPEAKER_00]: It's extremely costly, so this is something that I really want to improve with established feedback sessions which are one on one every week, which tried to be as honest as possible and thereby directional, and just constantly asking ourselves questions on both sides, you know, why does this make sense?

17:16 [SPEAKER_00]: Why does it make sense to continue working together? [SPEAKER_00]: what are we accomplishing, what is the other person's goal, what is my goal if I'm managing them, so that we're on top of the situation and we're sure that things are working because when they don't, it's very costly. [SPEAKER_00]: We've met a lot of other mistakes and, like I said, continue making. [SPEAKER_00]: I fund-raced in a market that was very frothy and I've done it and walked into it super

17:48 [SPEAKER_00]: And then things changed extremely quickly, and that fundraising cost the company, the product, and myself a ton of time. [SPEAKER_00]: Just because I wasn't well prepared and I thought it was going to be easy. [SPEAKER_00]: It turns out nothing is easy, and being all prepared matters is always good. [SPEAKER_00]: And on the product side, I mean, we've prioritized a lot of small customers, we're happy because we've learned, but sometimes we really miscalculated our pace of learning and the improvement of the product versus the impact it was having on the whole company, the time it was taking and how unique those cases and features were versus what we needed to do as a whole.

18:33 [SPEAKER_03]: This will be fun to ask. [SPEAKER_03]: So, what does the future look like for your product and for the team? [SPEAKER_00]: We just landed in a really exciting space from all perspectives. [SPEAKER_00]: When we started building, I think it was still a little bit early because [SPEAKER_00]: The world was excited about financial data being available for builders, but it hadn't quite caught on yet about how good the quality needs to be and that it wasn't there. [SPEAKER_00]: Right now, we are in a world where a...

19:07 [SPEAKER_00]: financial data seems like is going to be a part of not just every financial product, but also other software companies and products. [SPEAKER_00]: As we're seeing, even Shopify store dashes, apples of the world are becoming in a way financial services companies are starting to embed financial services, which means that impact of the state is going to be way beyond the traditional banks

19:37 [SPEAKER_00]: another massive thing that is super important for us as a company is that if before only people who could really use this data and access it were engineers, data scientists, and you needed that skill to be able to make the most of the data even do like simple queries and answer questions about your customers, [SPEAKER_00]: looking at this information. [SPEAKER_00]: Right now, there's a lot of capability with foundational language models and sort of the advances in that field to allow anybody to use our core and query it, build models from it, do forecasts, build recommendation engines, do whatever else whereas before you just need it, engineering skills.

20:21 [SPEAKER_00]: And like I said today, although there's a lot of

20:27 [SPEAKER_00]: There's nearly not enough of them and there should be more. [SPEAKER_00]: And I think there's going to be a big wave of people who will be given more tools to be able to leverage the ability to code without necessarily getting super deep into programming and they can be users or data. [SPEAKER_00]: Any knowledge worker powered by numbers and financial information is a 10x knowledge worker, whether it is in marketing, a product, investment, finance, underwriting and so on, so forth.

20:59 [SPEAKER_03]: Alright, let's switch to you. [SPEAKER_03]: Who influences the way that you work? [SPEAKER_03]: Name a person or many persons or something you look up to and what? [SPEAKER_00]: I consume a lot of information and sometimes way too much. [SPEAKER_00]: I need to have the discipline of stopping and distilling of what's coming from the outside. [SPEAKER_00]: And then what is real, what I think is real, what is important to me and block the noise, but also take the learnings from all the information that I consume.

21:29 [SPEAKER_00]: So through the years, I've become better and better at this. [SPEAKER_00]: And one of the people who really helps is, actually make a funder, who I started the company with, he is the opposite of me. [SPEAKER_00]: he's hyper focused, very narrowly focused at times. [SPEAKER_00]: He can do one task at a time and he will get it done no matter what. [SPEAKER_00]: He's played a major role in how I think about things as well and he is like the opposite mirror that I can look into and then adjust and improve.

22:01 [SPEAKER_00]: Although sometimes it's tough because you have to see your vices and accept them. [SPEAKER_00]: And then I've had a baby 18 months ago while I was fundraising for a round and running the company at the same time. [SPEAKER_00]: It was a very weird and interesting experience and today just seeing a new human growing up learning about the world and the way they do things gives me a lot of perspective. [SPEAKER_00]: probably anybody who has kids can relate to this and even if you don't.

22:28 [SPEAKER_00]: It's commonly repeated, but it's fascinating to go through that experience and apply it to other things. [SPEAKER_00]: I've suddenly discovered this expanded sort of capacity, expanded mindset, expanded patience for things, which I didn't have before, but it influences the way I work. [SPEAKER_03]: Last question, so you're getting on a plane, and you're sitting next to a young entrepreneur who's built the next big thing. [SPEAKER_03]: They're jazzed about it, but can't wait to show it off to the world, can we show it off to you right there on the plane?

22:57 [SPEAKER_03]: What advice do you give that person having gone down this road a bit? [SPEAKER_00]: Well first of all, I love people who build so I'm excited to see what they've built. [SPEAKER_00]: I think it's really important if you've built something you're very excited about to also be extremely open to what people are actually saying versus what you want to hear. [SPEAKER_00]: This is hard, you have to balance your dream and things that you just have to be delusional about in order to make something happen, otherwise it just doesn't happen, but at the same time being able to hear what people are saying and what they need and what matters to them.

23:38 [SPEAKER_00]: often the vision we have about the world is true, however it can be biased in so many different ways and to distill that into an actual product you need to be able to listen. [SPEAKER_00]: I still find it hard myself because I have so many assumptions and just having the courage to question them all the time without getting pessimistic and giving up this tough you need to build up a character

24:05 [SPEAKER_00]: I would definitely recommend somebody to start building up the user stories and to treat anything in any person in any conversation that they come across as a user story and to use the stories to create the bigger narrative that they want to see. [SPEAKER_03]: That's fantastic advice. [SPEAKER_03]: Well, Nari, thank you for being on the show today and thank you for telling the creation story of entropy. [SPEAKER_00]: Again, thank you for having me with the pleasure.

24:38 [SPEAKER_03]: Code Story is hosted and produced by Noah Labhart. [SPEAKER_03]: Be sure to subscribe on Apple Podcasts, Spotify or the podcasting app at your choice. [SPEAKER_03]: And when you get a chance, leave us a review. [SPEAKER_03]: Both things help us out tremendously.

24:54 [SPEAKER_03]: And thanks again for listening.