How RedStone is Connecting Financial AI Agents to Verifiable Data

[00:00:00] - [Speaker 0]
Your agents aren't producing accurate answers because they don't have a complete semantic understanding of your data, and Denodo is solving this and solving it through semantic consistency. Through semantic consistency, your agents can start making accurate predictions in real time. So see what else Denodo can do by visiting denodo.com to learn more. But now, let me introduce you to today's guest. Welcome back to the Tech Talks daily podcast.

[00:00:36] - [Speaker 0]
I have the cofounder of Redstone joining me on the podcast today, and we're gonna talk about something incredibly complex but in a language that everyone can understand. We're gonna talk about why smarter AI models won't solve the problem without accurate verifiable data and how infrastructure built to protect billions of dollars in in blockchain transactions is now also being used to keep AI agents from making disastrous mistakes. And if we have time, also look at what businesses should be doing before trusting autonomous systems with real money. We've got a lot to get through today. So without further ado, I'm gonna officially introduce you to him right now.

[00:01:20] - [Speaker 0]
So thank you for joining me on the podcast today. Can you tell everyone listening a little about who you are and what you do?

[00:01:27] - [Speaker 1]
I'm Marcin, cofounder of Redstone Oracle's and Credora Ratings. I've been building in the blockchain and crypto space since 2020, so over six years. We created one of the largest blockchain Oracle network with Redstone. We are focusing on making sure that the data that is delivered and utilized in the DeFi and on chain finance market at large and tokenized finance is reliable, accurate, and, for sure, diversified to the users that need that in those kind of environments. Previously, I was working at Google Cloud, had two startups that failed, that gave me a lot of experience, and wrote my bachelor thesis about blockchain and Ethereum in early days.

[00:02:14] - [Speaker 1]
So I've been in that sector for quite a while.

[00:02:18] - [Speaker 0]
Love it. And I love how you mentioned data right off the bat there because I think so many discussions that we see on our news feeds and podcasts and videos, etcetera, they all focus on models getting smarter. But you've argued that the real challenge is actually the data quality. So why has reliable, verifiable data become such an an important foundation for AI systems, particularly in financial services? Or has it always been important, but people kinda skipped it?

[00:02:47] - [Speaker 0]
What are you seeing here?

[00:02:49] - [Speaker 1]
Yeah. That's a really good question. Maybe let me take a step back because Redstone predominantly works in the blockchain sphere. That's where we started. And now we see this AI revolution coming in and actually tapping into the advantages and opportunities that tokenized assets and DeFi gives in general.

[00:03:10] - [Speaker 1]
So first, the data in the crypto space, it has to be accurate because financial transactions are usually of high value. Like, you're doing a swap or like any financial activity with millions of billions of dollars. Even a small percentage of mistake on the price or the data delivered can have catastrophic results, and that's the same case in the AI at large. An Oracle is a system that aggregates information from multiple sources. In our case, it's over 400 sources, and make sure that the final information delivered, for example, a price of an asset, liquidity, volatility, market cap, or any other financial information, is the most accurate according to as many sources as possible.

[00:03:59] - [Speaker 1]
Many of the times, it means, for example, a median out of 200 sources to make sure that the data you deliver is extremely robust and no one can skew that, or a dedicated statistics like a weighted average and so on and so forth. In AI realm, the importance of data is even more important because you have a muscle that is extremely extremely powerful, but if it's going to operate with bad, let's say, inputs, it's going to make disastrous results, especially that AI agents can be extremely creative when it comes to the paths they're taking going forward. So with financial infrastructure going twenty four seven and fully global with the tokenized rails and blockchain, because this is like one of the advantages that this industry gives, This accuracy of information about, for example, FX, so the exchange between currencies, or the accurate pricing in the off market hours, so when the stock markets are, for example, closed, is extremely important because agents can make a lot of hallucination when executing transactions.

[00:05:15] - [Speaker 0]
And for people listening and hearing about this side of the business for the first time, can you tell me a bit more about what Oracle does in the blockchain world? Because they might be unfamiliar with the the concept, why the technology has become so important for connecting digital systems with real world information, and and may have preconceptions about crypto, for example. So tell me more about that and the focus there at Oracle.

[00:05:40] - [Speaker 1]
Of course. So in blockchain systems, you have concepts such as smart contract, which is essentially a program that's executing based on an input. You can imagine a simple if, and that input has to be as accurate as possible, similar as we've discussed with AA Systems, but with smart contracts, usually, you have at stake enormous amount of cash, like, for example, billions of dollars. For example, Redstone is securing right now over 10,000,000,000 in total value locked, meaning that if our system goes down right now, $10,000,000,000 is at risk for people to lose or being stolen. Therefore, Oracle's role is to ensure that the data is delivered continuously.

[00:06:31] - [Speaker 1]
So this reliability piece is extremely important. We, as Redstone, for the past six years, have never had a single mispricing or daunting event, even when Yahoo Finance, AWS, or any other large organizations had shorter outrages, and then people remember them vividly because they, for example, cannot access a specific account or, like, say, Facebook is down for some time because one of the parties in their chain is not responding. For our business, it's utmost important that we are always upkeeping the infrastructure and delivering the information because those financial transactions are happening in real time and executing based on the data that we deliver. So an Oracle is a system that gets data from multiple places, aggregates, cleans that, and delivers to this smart contract for execution. Then when you have tokenized assets, for example, on those blockchains, Ethereum, Solana, and so on, you can create for AA agents those environments to operate with digital assets at large within twenty four seven realm.

[00:07:43] - [Speaker 0]
And Redstone was originally built to solve data challenges for smart contracts. So how are some of those same capabilities now being applied to AI agents? And do you see any similarities between the two worlds?

[00:07:57] - [Speaker 1]
Of course. I do see a lot of similarities, though I have to admit that AI agent space is extremely fast in the sense of, like, expanding. In the blockchain industry, evolution was happening, I would say, every six months. In AI, it's probably, like, even a month that you see something that is progressive progressing extremely fast. What we can see in terms of similar patterns in the two words is, one, the growth of interest in argentic finance, because this is where a lot of optimization can come into place, especially for users that are not so savvy utilizing, for example, wallets or other technologies within the blockchain sphere that are not so easy to comprehend and get used to, like, right away.

[00:08:53] - [Speaker 1]
They would prefer an interface that is familiar to them, for example, like a chatbot, and then they want to abstract away usage of the blockchain itself. So in both of those environments, the data delivered has to be exactly the same. The AA agent has to have, like, a particular interface to aggregate the data and take decisions based on that, and the on chain environment, it has to be on chain call. So an Oracle is usually we, for example, are pushing the data to the destination blockchain within specific interval. So this synchronization between the two words, for example, we deliver the data every five seconds and an AA agent is taking decisions every second, there is like a discrepancy between the two.

[00:09:42] - [Speaker 1]
So many of the times orchestration and making sure they're both in at least as perfect sync as possible becomes extremely extremely important. Then the second aspect is hallucination. This is one aspect this is one thing that we have not seen in on chain finance because there everything is deterministic in terms of smart contracts. In AI world, like the to prevent from hallucinations, have to make sure that the outlier detection is even more rigid and ensuring that the information you feed the agent itself will not skew the perception of a market to take a particular directional trade.

[00:10:30] - [Speaker 0]
And I think this year, we're we're hearing more and more about autonomous AI agents, how they can execute financial tasks in everything from payments, tradings, or even risk analysis. But what do you think needs to happen behind the scenes before the majority of organizations will will get that trust that an AI agent can make real time decisions involving real money? Because it feels like there is a slight trust issue there, and for good reason. But what needs to happen?

[00:11:00] - [Speaker 1]
I believe, coming back to what what I mentioned, this hallucination has to drop in terms of the severity and also, like, how often it happens. I found the statistics that AI hallucination contributed to estimated $2,000,000,000 in avoidable trading losses in q one of twenty twenty six alone. So within three months, 2,000,000,000 is naturally like a huge amount. And unguarded LLMs hallucinate usually on about 20% of financial tasks, because they are just usually not accustomed to those kind of areas. So I do believe in order for people to get more comfortable, this hit ratio of hallucination and disastrous decisions from agents just has have to drop.

[00:11:49] - [Speaker 1]
And to achieve that, the data piece is predominantly important because you can get better and better on the model itself, and this is what we are seeing across the board with Anthropeak, OpenAI, and every other organization dropping more efficient models themselves, but the data on which they are operating has to keep improving. It cannot be treated the same as a human human eye would see on a display or somewhere else. Right? It has to be very well positioned and understood in terms of, like, what can cause a hallucination, and then if it's repeatedly observed, take adjustments so that the final output of the information delivered is, yeah, particularly fine for the AI agent themselves.

[00:12:46] - [Speaker 0]
Yeah. And I think one of the reasons I asked that question is, obviously, even small errors in financial data can have major consequences. So are there any other safeguards that organizations or leaders listening should be putting in place to ensure that their AI systems are working with accurate, timely, and verifiable information rather than making decisions based on on flawed inputs. Anything else you'd like to add to that?

[00:13:11] - [Speaker 1]
Well, there's probably thousands of GARIOs you can input implement right now. Right? And then the problem is with chicken and egg. So if you're going to impose too much, then the value that the agent is actually delivering can be way smaller. I was discussing Glitch even today with a friend, an agent bot that was playing on prediction markets that throughout three months generated $200,000, which was, like, very impressive.

[00:13:39] - [Speaker 1]
And then in the context of two transactions, it lost 250. So it just showcases this kind of, like, disastrous effect if this hallucination is actually happening. That's one also the of the reasons that we, as Redstone, expanded project called Credora, which is risk rating, that assess risk related to interacting with particular financial assets so that AA agents can map out not only their potential gains or yield, which is the amount of money you can generate holding an asset throughout, for example, a twelve month period, but this should also put into account the risk that is associated with that particular interaction. So, for example, we give ratings between d to a to money market funds, private credit, reinsurance, or in general trading strategies. So that you can imagine an a agent that has limitless opportunities to invest, but you can instruct it, hey, operate only within the realm of the assets that are rated by Credora.

[00:14:51] - [Speaker 1]
And then it doesn't mean such an agent cannot take a riskier bet and decision on a high yielding product with a d, like with very low risk rating. But then this agent is just better informed that it's taking such a decision. Or you can instruct it to only operate within the Credora rated products that are b or higher. Right? Or similar other logic brackets.

[00:15:17] - [Speaker 1]
So in general, for professionals, I would say understand, like, what the agent's primary goal is and what can be the most severe or frequent hallucinations or problems, and then try to impose guardlays or solutions that are going to address specifically that. Because we are coming from the financial world, we are mainly looking at the financial risk aspect. Right? So this is what I would add.

[00:15:47] - [Speaker 0]
And, obviously, you've worked with both decentralized finance and more traditional financial institutions. And I think that gives you somewhat of a unique vantage point. So I'm curious. Are you are you seeing these two worlds beginning to converge when it comes to AI infrastructure? And if it is, are are there any lessons that maybe established enterprises could learn from the way that blockchain ecosystems approach things like trust and data verification?

[00:16:15] - [Speaker 0]
Any big lessons there?

[00:16:17] - [Speaker 1]
Yeah. Interestingly enough, I came back from New York this week after three weeks where I was talking with the largest financial institutions in the world. We went to the office of Nasdaq, NICEE, London Stock Exchange, WisdomTree, and many many other giants, learning about their current expansion to on chain finance and also, like, how they're utilizing AI agents. And one thing was very clear is all of them feel enormous pressure to innovate because they know the pace of innovation is only accelerating. So it's really refreshing to see that the approach is not like, hey, we are the largest and we don't have to move.

[00:17:00] - [Speaker 1]
We are going to be here for the next hundred years. No. Rather, it's we are seeing where the puck is going, and we want to also follow that naturally within our pace and time frame of the large organization. We organized an institutional conference called Tokenize This, three days of intense conversation with DTCC, Fidelity, Invesco, Franklin Templeton, and so on. And what they are the most interested about is the tokenization game, which has, in the past period, accelerated enormously.

[00:17:33] - [Speaker 1]
So tokenized real world assets grew in the last eighteen months from $5,000,000,000 to $33,000,000,000, where it sits today, and it's continued growing. And right now, when you have those tokenized assets, many of them are thinking, okay, what are the AA agents used by either assets organizations or, the general public users that could tap into those tokenized securities and other assets as a yield opportunity, increasing the distribution and availability of products the institutional players are offering to the broader market. And many of the times, it also means going outside of The US. Right? Because they think, okay, The US market is somewhat already penetrated, but they really want to distribute it even further across the globe because there is such a big demand for them.

[00:18:28] - [Speaker 1]
Another trend that I did see is that majority of those firms, and that I found statistic for 78% financial firms, have already deployed AI for data analysis. So they are first utilizing it for getting information about the market and how it's operating. And then the next step is going to take the best ones that are performing in lessons into the executional ones. So operating on the actual assets, and suggesting, for example, best allocations to either trading professionals or the general general public.

[00:19:07] - [Speaker 0]
And finally, for any business leader listening who is excited by the potential of AI agents, but do find themselves equally concerned about reliability and governance, etcetera, any practical advice that you would offer before they allow autonomous systems to take on financial decisions or even other high value business processes? Anything you'd want them to take away?

[00:19:30] - [Speaker 1]
I believe before taking, like, serious decisions, talk with people that have already implemented some of those systems. One thing that I learned is many of the times right now there are people within your network that have already interacted. So if you're feeling FOMO, you can ask around, like, even on Twitter, like, hey, has you ever deployed or on LinkedIn an AE agent that did x y z? So let's chat, and many of the people are going to come over and share their experiences. Second is, of course, first test in sandboxed and closed environments before going with real money or real use cases so that you can understand first what are the critical things that can go wrong and how you can implement any kind of response policies should there be an incident, like with an AA agent.

[00:20:25] - [Speaker 1]
And another thing I I believe will be learn more about tokenized assets. I mean, sorry I'm biased, but I do think this is the future of financial game. The decrease of cost, higher availability, twenty four seven markets, and essentially cutting out many of the middlemen's from the system is the perfect setup for AA agents to revolutionize at scale. Right? Because the financial stack as we know it today was built, like, incrementally, adding layers that are redone that are not necessary, and that are increasing the cost and the time needed for settlement transactions and many other operations.

[00:21:11] - [Speaker 1]
Whereas the blockchains, like Ethereum, Solana, Canto, Stellar and many others, fundamentally are like unified layer that allows for global finance to operate within seconds and at a fraction of the cost of those financial systems. So this setup is a perfect environment for agents to operate and innovate instead of, like, a guarded closed systems. And therefore, I think keeping track of what's happening in tokenized finance, DeFi, will allow users to also create most meaningful implementations with AI at large.

[00:21:50] - [Speaker 0]
Well, we covered so much today from Redstone, the fastest growing blockchain, Oracle, where they're delivering, I think, 400 plus price updates per second across 110 blockchains. Absolutely phenomenal to watching, to what you're doing now, watching AI and fintech teams deploy autonomous agents and, obviously, how Oracle's infrastructure is evolving into a foundational layer for AI development. So many big talking points for anybody listening that would like to, learn more about you, your work, and anything we discussed today. Where would you like me to point everyone?

[00:22:26] - [Speaker 1]
Yeah. So you can find me on Twitter, x at marcin redstone, and redstone underscore defi are our handles. I am very active on LinkedIn. Happy to answer any questions over there and chat with people that are interested in Oracle's tokenization or agents at large. And I will be present at many of the conferences this year as a speaker, Korea Blockchain Week in Seoul, Token twenty forty nine in Singapore, Solana Breakwind in London, and naturally tokenize this in New York City, the next edition.

[00:23:01] - [Speaker 1]
Hopefully, soon I'm going to come to London as well to meet some of the financial folks over there, and then Nail, for sure, we can grab a coffee and chat further about the tokenization and agents revolution.

[00:23:14] - [Speaker 0]
That sounds like a great deal to me, and I cannot thank you enough for coming on here and giving a very practical look at agent driven finance. Everything that you're seeing and what preparations can be put in place, and also put it all in a language that everybody can understand. So I'll include links to everything that you mentioned there, and I encourage anyone listening that's going to any of those events, please reach out to to you, and, hopefully, we can keep this conversation going. But more than anything, thank you for starting it today. Really appreciate your time.

[00:23:44] - [Speaker 1]
Brilliant to be over here. Thank you, Neil, and see you everyone soon. Cheers.

[00:23:48] - [Speaker 0]
I think today's conversation was a powerful reminder that autonomous finance will only be reliable as the information feeding it. And Marcin explained why AI agents need more than access to markets and the ability to execute transactions. What they actually need is trusted data, risk intelligence, clear limits, and testing environments. Environments where businesses can understand what might go wrong before any real money gets involved. And I also loved his advice for business leaders.

[00:24:23] - [Speaker 0]
They're feeling pressure to move quickly. Simply talk to people who have already deployed these systems, people that have already learned from their mistakes and tested in controlled environments before handing over the keys. But I'd love to hear your thoughts. Techtalksnetwork.com, how much financial autonomy are you comfortable with giving an AI agent, and what would it need to prove before you trusted it with your money? Let me know.

[00:24:51] - [Speaker 0]
That's it for today, though. I'll be back again tomorrow with another guest, but thanks for listening as always. Bye for now.