Interview with Caesar Sengupta

INCH: Tell me a bit about how Arta came to be.

SENGUPTA: We started up because of a pain point we faced in our own lives. If you think about most people climbing the professional ladder, once they accumulate a bit of wealth they make some investments here and there, and put everything in spreadsheets, then eventually they find out about private banks. That happened to a bunch of us, and we had a few big observations. First, there are a lot of financial superpowers that we didn't even know about. Forget about access, we didn't even know that they existed. But, at the same time, the barriers to entry were very high — most private banks won’t talk to you unless you're moving $10m or more in. And, the process was extremely manual and old school. Everything in our lives lived on our phones, except for our wealth.

We realized this is a problem faced by a large number of people around the world. If you look at people with anything from $100,000 to $5 million in wealth, in total they own about $270 trillion globally – it’s a huge market. We ended up figuring out the best way to solve this is to bring together a set of people who are deep in technology, a set of people who are deep in finance and make something that is easily understandable for members.

Q: Let’s talk about how you’re using AI. Your website lists AI-Managed Portfolios as a feature, could you talk more about those?

We’ve actually started calling them quant strategies here since some people were getting confused. They’re using traditional machine learning. Our flagship is Defensive Growth, a min-volatility strategy. If you want to invest in it, you can pick the risk level that you're comfortable with and customize it based on sector. So you can say, I work in tech, so I should be diversified away from tech, and now we will use machine learning to solve min-volatility for you and invest based on that.

Another one is what in Singapore we call S&P 499, in the U.S. it's called direct indexing. Here, instead of buying an S&P 500 portfolio, we buy the components of it and then buy and sell when things go up and down. That way you can maintain the exposure and the performance of the index, but generate tax losses. So we can generate like, 2 to 10% of tax losses while tracking S&P 500. That's very classic, machine learning, not generative AI-based stuff.

Q: Are you using any generative AI? And if so, are you finding any value with it?

Quite a lot, for example with asset analysis. When you go to a private bank and ask if you should buy Google, typically the banker will send you a bunch of analyst reports. What we've done is taken a bunch of public data analyst data, SEC filings, our own licensed data from Bloomberg, the Wall Street Journal, and a couple of other data sources, and then, based on all of that, we generate a very simple asset analysis daily.

Asset analysis of Alphabet Inc., generated by Arta AI Copilot

Q: This looks like a retrieval-augmented generation approach?

Exactly, yeah, it is RAG, but we use a multi-agent system. Where it's a public source, we will use commercial LLMs, but for our own proprietary licensed stuff we use a set of different Llama models (Llama is Meta’s open source LLM). Depending on the data source, these different co-pilots work on each data source, and it’s eventually composed together into one cohesive thing.

But it has to be a multi-agent system, because there’s some data we can’t send to commercial labs like openAI or Anthropic. We do use the commercial labs sometimes because they have much larger context windows (the maximum length of text you can send to an LLM). So if you're throwing a very large analyst report at it, you kind of need something with a larger context window.

Q: Could you talk a little about the process of taking that system from initial idea to final product?

It was several months of reading through the right architecture. First we had to figure out, what is the end user experience, right? What does the end user want? So we spend a bunch of time thinking about, what are the key use cases that actually resonate with the end user? I described our simplest use case, but if I show a few more then I can talk about the architecture.

Q: Sure, let’s check out some more features.

For the second feature that we wanted to do, the question was how we take something that a private bank user has and use AI to make it available to many more people. One typical thing that a private banker would do with their client every quarter is sit down and tell them “here's how your portfolio is doing, let me contextualize it to the market and give you some more analysis.” We can do exactly the same thing, but now using AI we'll generate a portfolio recap. It tells you how your account is doing, what within your account has performed the best, and then compares it to the market. It also compares it to an MSCI risk model that we license. So you can find out, for example, how interest rates have affected your investments. We also show a set of news summaries that are based on the specific assets inside your portfolio. That’s not just at the ETF level, but deep inside, considering what's contained in the ETF.

There’s one other feature here, and you will laugh at this, because it's super simple from an LLM perspective to do, but it blows everybody in the room away when we demo it. The default text we generate is fairly sophisticated, but we can use an LLM to turn it into easy reading. Or in Singapore, there's a dialect called Singlish, so we can turn it into Singlish. This has turned out to be one of the most popular features we have. Ultimately what we’re getting to is an AI banker that talks to you, that understands you and that talks at your level, it doesn't talk down to you. And these are the places where we think LLMs will have the biggest impact, helping to make finance much more approachable.

Arta’s AI Portfolio Recap summarizes performance over a given time period, in a style of the user’s choice.

Where we’re going eventually with that is ultimately a voice-based AI banker, which we haven’t launched yet but can demo.

(At this point Sengupta pulls out his phone and asks their co-pilot how his portfolio is doing. After a few seconds, the assistant launches into a brief summary of overall performance, and mentions some individual high-performing ETFs.)

Again this is a multi-agent system, using a set of different labs’ models. This co-pilot is in my account so it's got access to my data in a private, secure manner. It also has access to all of our products, so I can ask about private credit and we have a co-pilot that looks at Arta’s stuff. Or if I'd asked it, “how expensive is U.S. vs U.K. education if I want to send my kids there”, then a different co-pilot would go look at the web, and bring that back to answer the question.

Q: So when you say multi-agent, you’ve got one LLM that’s interacting with the user and using function calling, giving it tools like web search?

Right. There's a bunch of function calling, and then there's a whole sequence of agents that are getting called, as well a supervisor model that makes sure it all comes together. The voice stuff won’t ship for a few months, but everything else we’re showing has shipped in Singapore, where you can access it globally.

Q: And there’s another feature where you’re leveraging GenAI?

We have a concept called satellite portfolios, where if you have a big belief in a particular type of investment we can create something like a mini-ETF for you. For example, if you want to invest in the Magnificent Seven, one way is to just buy the seven stocks, or a more structured way is us buying them on your behalf and keeping them balanced, so that you're always allocated in that ratio. That’s customizable: If you want to remove Tesla from your Magnificent Seven ETF, you can do it with a click. You can also check the performance of your allocation over the last year or three years. There's no AI in this, this is hardcore finance stuff, but the reason I'm talking about this is because when you're doing AI and investing, you have to put the hardcore finance stuff and the LLM stuff together. So now we can ask how we make this better using LLMs: letting a user create a satellite portfolio using LLMs.

What it does is first take the overall theme and use reasoning to break it into sub themes, and each of these kicks off a separate co-pilot. Each co-pilot has access to different data sources based on whether they're public, proprietary, or we’ve constructed the data source ourselves. Next, a separate master LLM checks if each of these investment copilots has done their work correctly. It's exactly the way a hedge fund would do it with analysts, right? They would send an analyst off to research a particular data stream, which is how we've comprised the whole thing. Once all of this comes back, it gets put into one consolidated investment memo. For example, for the Nvidia supply chain, it would understand things like specialty gas and material providers that will give you exposure. For the user we'll describe what the steps the agents went through, as well as give people links to the resources the agents used, so they can go consult these resources to validate it.

Arta’s AI Co-pilot generates a thematic investing basket using multiple agents in parallel

Q: When you're implementing these features using AI, what is the balance of time spent on the AI side of that feature versus the UX side of it?

It’s a hard question to answer, but we deeply care about usability. AI can be quite intimidating at times. Especially in the investment world, being able to understand what the AI is doing is super important, that's why we've spent a lot of time thinking through the UX. I don't know if I can say 60% on UX versus 40% on AI, but all I can say is we've spent a huge amount of time on it. They have to work very tightly together.

Sharing the resources the AI copilot used to build a portfolio is part of the match of AI and UX, because we are in the trust business, right? The AI has to invoke trust, and the best way to do that is to show a lot of detail. The second level of that is showing explanations. For example, in a portfolio covering Nvidia’s supply chain, you could check why the AI selected HP, and we'll tell you it’s because the AI found out HP integrates Nvidia’s products. If you like that answer, you can keep it, and if you don't like that answer, you can remove HP. We can then go and see how this strategy would have performed over the last year, comparing it to the S&P 500. Again, it’s very important to corroborate and substantiate the AI. It adds validity to the AI and makes it much more usable.

Q: With this ETF generation, you're stringing along multiple LLMs. Did you run into issues with hallucinations or reliability? Or did you find it worked well out of the box?

We've tried to reduce hallucinations by figuring out which data sources to use for what, as well as fine-tuning our internal Llamas with additional data. But the biggest part of safety comes with actually presenting the user with sources and hard financial data. At the end of the day, the combination of the two is what makes the AI actually usable and powerful. The AI can generate a report which you can read, but you still have hard financial data that helps you evaluate whether this strategy would do well or not. That ground truth is super important.

Those are the AI features that we've gotten going so far. We have a couple of bank partners that are looking at taking the asset summary and the portfolio recap, as well as a couple of asset managers that are interested in using this like an analyst. Instead of having a team run through and come up with strategies taking weeks, it'll help them do this in a few days, a few hours, a few minutes.

Q: You’re working on launching a B2B offering for banks – tell me more about that.

A: There are essentially 3 ways we work with banks. First is Arta Embedded: for a retail bank they can offer investment options powered by Arta, like a public market portfolio strategy or a structured product. We have a white label offering, where they take the whole product in their brand and ship it in a new country. And we also have AI copilots that are largely available as APIs. We’re offering every feature in our platform as an API, and a couple of banks are already integrating into that. Some are actually looking at it as a tool for their private bankers – for example with portfolio recap.

Q: Some large systemic banks tell us that it's difficult to work with startups because banking is so regulated and they operate in many different jurisdictions. Do you have to make any concessions to work with your partners, and do they need to make concessions to work with you?

I think it's give and take on both sides. We've typically found more digital banks are moving faster. In larger banks where there’s strong alignment at a senior level we see a lot of interest. For example, there are a couple of European banks where they see the vision at the wealth head-level or the CEO-level, but there are detailed challenges on both sides. Ultimately though, we're complementary. I don't want to have 1000s of employees in Germany, and this kind of technology is going to take the banks a long time to build. In the longer run, the B2B2C business is very interesting for us, because many of the banks have such a strong understanding of their market. For us it's actually very easy to change most of this stuff technically, but the challenge is understanding that market, understanding the user, and understanding the regulations.

Q: You're based in Singapore, and you’ve released all your features in Singapore. What do you think of Singapore as a hub for AI innovation and finance?

We actually have two teams, based in Singapore and Silicon Valley, but we launched a lot of the AI features starting with Singapore. The government here has been very pro-AI and very eager for us to show how we can push the boundaries with AI. Earlier this year, the Economic Development Board of Singapore made an investment in us, a strategic investment, primarily with a view to advancing AI in wealth. From a governmental and regulatory point of view, Singapore is a very good market for us to push the boundaries and really explore how AI gets done. But the teams are building it across both offices.

Q: Do you have a horizon with regards to Europe? Do you think you'll be releasing in Europe soon, or is there a lot of additional work you'd need to do to make that work.

It depends on the regulators. In the next few weeks, we'll release some of these features in the US. Those are the two markets that we've opened, and in every other market we work with a partner bank. Singapore lets us onboard people globally, so if somebody comes to us directly, they can use all of our features. But to launch in a particular European market, we would have to work with a local partner there, and then we'd have to work through the regulations and all of the specific issues that come up around that.

Q: Some employees at banks are concerned about accelerating AI adoption, maybe regarding job replacement or task automation. Do you have any thoughts on what business leaders can do to keep teams on-side?

I think that's a great question. Today most banks want to serve a lot more clients. Most RMs (relationship managers) would gladly take on 10x more if they could, but they spend a lot of time dealing with very basic stuff. My sense is that really progressive leaders are helping their teams understand that AI can actually help them serve more people and expand the business a lot more. The vast majority of people today don't have access to wealth management, and most banks are raising the minimum levels because they don't have a way to solve for it. There are also markets across the world where the number of wealth managers is reducing. So I think if you have a good wealth manager or a financial advisor, you can supercharge them with AI. That's a very compelling story. That's my approach of how to think about AI, at least in our field. There might be other fields where AI is much more threatening to jobs in the short run, but I think in the wealth space, there's just so much opportunity for expansion and democratization that AI is just going to make it more possible for people to get high quality services.

Today in Singapore, I hear about private banks where even if you have $10 million in, they want to know that you actually have $25 million in total. If they can go to around a million, suddenly many more people can get started. So I think that's the path, and that's what a lot of our partner banks are interested in.

Q: We’ve started to see bank leaders grapple with the costs of generative AI. You're running multiple agents, each making lots of calls to LLMs. Is the cost of generative AI a concern?

Oh yes, it is. But as a startup, we’re focused first on trying to figure out whether we can add value to the user. If we can add value for the user, then I'm sure revenue will follow. We are at that stage where we've built some cool technology, and now we are trying to see who ends up using it.