Autonomous On-Chain Intelligence: How AI Agents Will Transform Web3 User Behavior & Payments

Transcript

Hey everyone. Thank you so much for coming out today. Con is quite pretty, so I know there's a lot of other things you could be doing, but I I appreciate your time. So today I'm going to discuss how the AI behind Blocksite is changing payments with blockchain as the execution layer. Um we have three decades of NASA-funded predictive modeling and that's the brain behind our agent that suggests engages and guides users seamlessly on chain. So what you've probably been asking yourself since the last side is what is Blocksite?

Here is a photo of our dashboard that a client would receive. So it's AI-powered analytics that transforms raw user data into actionable intelligence for e-commerce. And the output for our clients is that they understand their users. The overall goal is to reduce churn. As we all know, this is a big thing in blockchain with applications losing users relatively quickly. Um both on the front end, we help guide them and through on chain actions. So you can see that there's some strategies you can implement like retaining users through airdrops, one-on-one outreach. Right now it's through call data, but we want to set up a chat system as well. Um

and sending coupons through our payments infrastructure. We're not just an analytics tool. We're built on NASA-grade AI. And this is a predictive intelligence engine that's purpose-built That's what Blocksite is. So these are the founders. Um two of them are not here today. I'll start by I'll just go in order. So Stefano Vercesi, he has 10 years of software experience. He's bootstrapped three products into including Coinrule as a lead developer. Um, he's built a lot of smart contracts. He worked for Builders DAO and Encode.

Um, and he has 4 years of deep blockchain experience. This is me. My name is Devin Martins. Maybe should have introduced myself at the beginning, but nice to meet you all. Um, I've architected AI-driven trading engines. So, I've worked for market makers, um, and managed over 50 million in liquidity. I led the build for Studio Chain. It's a layer two that running on Arbitrum Nitro. Um, both the node infrastructure and the uh the contracts that run underneath that allow users to stake and help run the network.

I worked on a layer one called Haven One for institutions, um, a set of smart contracts built on top of Quorum to help make a more safe chain so institutions can transact without a lot of the risk. And uh I did all the My Pet Hooligan contracts minus their original just basic NFT contracts, um the Hooligans and the Zuck Bots. So, all of the gaming infrastructure contracts and their marketplace. And the USA National Hockey League's Breakaway contracts, the pack opening experience. Uh Dr. Petrus Martins has over 30 years of NASA and NSF funded research in machine learning and

predictive modeling. He's contributed to Artemis, um specifically around astronaut safety because he does research on the sun uh and uses AI to predict solar flares. He's currently a professor at Georgia State University. And uh he's also taught at Harvard and Montana State University. Go Bobcats. I'm going to give him a little bit of credit since I'm talking a lot about his work today. He has over 400 publications. Uh he almost 5,200 citations across machine learning research, over 51,000 researchgate reads, and as I said before, 30 years of unbroken grant funding from NASA. He's worked for NASA

itself and ESA. And this long title, what it means here is he advises Congress directly on solar flares and helps them really map out like uh what the military does based on this. So, what it's called NASA Advisory Council on Aeronautics. He's the appointed advisor. And he's worked on the SOHO mission. He oversaw 12 scientific instruments aboard well, SOHO landmarks so solar observatory.

So, where we're at as a company right now. This is our payment widget. It is a simple script tag that you can drop in to any uh phone app or uh website. Just an HTML tag. It's very easy to use for an integrate. We have one end-to-end integration right now. Um Carrot Shop, it's a gift card store because I cannot ship things from my house. It would never work out. Um and it's a marketplace for My Pet Hooligan. It can be custom-made as you can see uh there's um My Pet Hooligan wearing a Santa hat.

Um it and the partner behind My Pet Hooligan is MGI Studios, the Carrot Ecosystem. They're Animoca Brands backed. Um they've been quite successful in the gaming industry for blockchain. We accept payments on any EVM chain. Um we detect the transactions in real time, and we return suggestions based on the strategy of our client. So, this is the first pipeline for our um AI end-to-end payments. This is why it's important to us. This is why I showed you guys this slide today.

So, the space weather analytics for solar flares, I'm going to say that once throughout this presentation cuz it's a lot, is our benchmark data set. So, it has 51 predictive parameters. This is important because um well, these are the our input features for all the machine learning models. Over 8 years this data was collected and it covers 10,000 solar flares. So, it's been co-created by Martin's team at GSU and is hosted on the Harvard Dataverse. That's the gold standard for machine learning and AI.

Um every model Block site uses is validated against this data set. It has strong architecture, multivariate time series rolling windows, and it let it leverages stratified sampling. This is something I'll explain later in this talk because I'm going to compare the two AI models. We have different features. Obviously, blockchain data is much different, but we're looking at wallet analytics um to replace the active regions and the medic the the method but the methodology is almost identical.

So, we have four classifiers. These are machine learning classifiers. The first here is the support vector machine. And what it does here, I've added pictures to try to explain all of um our AI work. So, we find the best the hyperplane that best separates two classes in the feature space. So, for Artemis, this is a 51 uh dimensional feature space between the classes that will flare, the regions of the sun that will flare, and that will not flare.

It segments wallets wallets between power users at risk organic versus bots and retain versus churning for BlockSeite. Then we have a decision tree. And Nano Banana made this nice picture. And what it we're explaining, right, is the root of the decision tree is our first question. There are several ways you can um classify each branch and how it works. The first is a mathematical equation, right? So, let's say balance is dropped by 40%.

Then we make a decision based on this. Um in terms of the Artemis mission, we're looking at the helicity. If it's over a certain percent, this is an example. And if it's over we're looking at something that's over a current density, we're going to predict a flare based on this. So multi-layer perceptron, this is a neural network. It's a feed-forward neural network, meaning the movement goes forward from input to output.

Something I'm going to discuss a little bit is what's going on in these hidden layers when we talk about XAI. But this is the input data. These are your features. These are your parameters. And this is the output. This is the prediction. So, what we're looking for for Artemis is we're discovering non-linear compound relationships across all parameters simultaneously. Work in BlockSeite. This captures compound wallet signals that are invisible to simple rules. So, let's say low governance interaction, for example, rising gas spends in prices. And then we're going to look at new protocol activity. So, is your wallet engaging in a similar

protocol to yours? This is a way to predict early churn. Our next classifier is a graph neural network. This is really important because we can look at nodes which are these dots here, right? This is again input hidden output. Um and we can see what the data points are, but we also want to see how they relate, especially in blockchain. This is important because a bunch of wallets, they can look normal alone, right?

But if we look at them together, let's say they maybe they're funded from the same source, maybe they're doing actions in synchrony. And maybe they're draining liquidity from your app. That's farming, and this is very common. So this is something we want to catch early and engage in. Our agent if you allow it to do so, can cut the yield on these wallets. So your protocol isn't getting farmed by bots. And for the sun, the magnetic field regions, they don't just have properties, right? Different regions of the sun are all still part of the sun, and they connect together.

So just to go through what we're doing from start to finish, the mission, the output, what I discussed earlier, now that we have a little bit more context. For Artemis, images were taken every 12 minutes. For Block site, on-chain transactions, um and block block level data. We're looking at the 51 sharp magnetic parameters, wallet behavioral features. Our uh classifiers, support vector machines, decision trees, multi-layer perceptrons, graph neural networks, and a few more that we'll talk about at another talk, but some classify but these classifiers,

they're running on wallet streams, same classifiers. Now we're going to talk about XAI here because this is the explain portion. This is important. You want to understand why our users are doing what they're doing. And for Artemis, these the goal is that the astronauts get a warning 24 to 72 hours in advance rather than 8 seconds because solar flares travel at light speed and when you're outside of the Earth magnetosphere, 8 seconds isn't going to save you. So just to give a quick overview, they will actually go in the craft and pack stuff around them to protect themselves or a lunar evacuation, depending.

Um so and for Blocksite, we have real-time churn and growth alerts. So you can interact with a growing wallet or a churning wallet. So explainable AI is what I'm going to talk about next. Somebody asked me during this conference if it had something to do with Twitter. It doesn't. XAI is why the system said what it said. And for a system that could trigger a lunar evacuation or flag your top whale as a churn risk if you hold a token, for example.

The black box that has historically been what we've used for neural networks saying, "Okay, here's your prediction." Not really knowing how to use it. This is why XAI is trending right now. I'm sure you've heard the word before. It's not acceptable. We want to explain every decision. So um This is a I know you all recognize this. This is a soccer field and it's because I'm going to use it for an analogy. Shapley additive explanations distribute prediction credit across every feature using game theory telling you exactly how much each input each variable contributed.

This analogy, well, we're looking at soccer players. Let's say there's a win or a loss. We have a churn loss or we have a a bunch of new whale emergence. We're going to this quantifies which wallet behaviors drove this. So, governance drop governance interaction. And it's not just based on governance, but this is an example I've used throughout because I knew the crowd here at ECC would understand. What drove the alert? IG non-Instagram integrated gradients.

This is specifically used on our neural network models. So, multi-layer perceptrons and graph neural networks. It's a really powerful XAI technique. After it attributes a model prediction back to its input features measuring how much each feature contributed along the path from baseline input. So, think of it as we're looking back to the baseline of when this trajectory started. When your churn began of your users, when the when all of a sudden your protocol became more interesting.

This does a few things. It not only tells you more about the features, but it also can help you make relationships between what's going on off-chain, what your company is doing, and what is happening on-chain with your users, how they're responding to you. So, yeah, pinpoints the exact moment and metric where a wallet began its trajectory. All right. Now Grad-CAM. How it works, right? It highlights spatial regions of the model focused on when it's making its call.

So, if you think of this as a neural network, and these are your wallets, we're showing which wallet costers or on-chain geo zones drove the predictions, and this makes your data much more interpretable. So, how are we doing this? Because there are things that are more important than other things. Cost-sensitive learning. So, for Artemis, we're going to penalize a miss flare prediction for very obvious reasons more than false alarms. And for BlockSci, we're going to penalize miss miss churn signals for your most high-profile wallets more than false positives.

However, the ideal scenario is perfection right? Stratified sampling. So, we're partitioning the data set into So, each training fold contains an equal number of flare events. I've done this with soccer teams, sorting them by teams. So, we have the same amount in each training set. This ensures churn events and whale emergence emergence events are equally represented in your training data, and we're not training our model on something without these patterns.

All right. Nana Banana gave me this picture, and I'm going to explain it to you. CGAN synthesis or conditional GANs is where you have a discriminator and a generator, and we picture them like as fighting each other. The generator is generating fake data that looks real, and the discriminator is catching it. And they're getting better as they keep going through this pattern. So, we For us, we're looking at um synthetic churn trajectories and whale emergence. For Artemis, we're looking at realistic but fake solar you know, solar flare trajectories. So, to summarize this, your unit of measure is an active region of the sun. Ours is a wallet. 51 magnetic parameters on

chain metrics that I don't want to give the exact number for. These were sampled every 12 minutes for SWAN SF. We sampled it at BlockChat at a block level, so we can get instant feedback or at an hourly level depending on our models because they work together. The question is where is this region's trajectory headed? The question we're answering for our applications is is this wallet headed toward turn growth or anomaly? So, I must have done this accidentally this morning and slid the bar a little up. I apologize. Smarter suggestions, deeper engagements, frictionless settlement.

That's the blockchain. The blockchain is handling the trust and the payments and the agent is handling the user experience in ways we weren't able to do so before. We ingest continuous multivariate time series data across millions of on-chain entities. Rare models or model models that I'm getting a little nervous here. We model rare high consequence events with early detection and deliver explainable outputs. Dr. Petras has spent 30 years building this exact machinery and Blocksite is where it's meeting blockchain now.

So, we have space grade AI e-commerce intelligence. Blocksite isn't just a dashboard. It's a predictive intelligence system built on three decades of NASA funded AI research that now applies to the most data rich frontier in technology. Just like every other company uses to advertise in web two and maintain users, we're using it on user data without without giving away your dog's name and all these locations like the web two companies are. So, the same AI algorithms and so that save astronauts' lives, support vector machines, decision trees, multi-layer perceptrons, graph neural networks,

Shap, Grad-CAM, and on this new new domain where the data is really rich on chain analytics for protocols, payments, and Web3 products. Please feel free to come and ask me questions afterward and thank you very much ECC for having me.