BlockSight | How On-Chain AI Sees the Move Before It Happens - Devon Martens | Common S3nse - 2026

[music] All right. So, hi everyone. I'm Devon Martins. I am the CEO of Block Site Labs. I also represented uh our lead um investor, Panther Hollow.

Uh the investor has come on as our chief um strategy officer. So, we're kind of helping each other's businesses there. So, that was why I was judging yesterday. Um, however, my full-time gig is uh Blockite. So, what Blockite does is we take data [snorts] that's already on chain.

Yes, I'm aware this is a privacy con um conference. So, let me just before you know everyone gets upset with me, start by saying that we work with privacy chains. We are on the Canton accelerator program right now. However, whether or not you disagree or agree with priv privacy, we agree that things should be private. To a degree, your competitors shouldn't have all your sales data, etc.

That's fair. But there are regulations that you need to follow and they're always going to be there. And that data needs to be delivered to the right places. And in order to stop people from doing things you don't want to do them to do or that you don't want to do. Hopefully, you don't want to do them on chain.

You need to be aware of the data. So, we know that user retention is very very difficult on chain. We're looking at a 93% turnover. We're looking at, you know, 42% of liquidity leaving right after launch, 70% within three days. So, liquidity providers that are usually being paid for the service.

They need to be held accountable. You can't dump on a project right after it starts if they are paying you to add value. And then of course we all know about the Cal Dow hack. And one of our, you know, main products is sort of a circuit breaker type deal that stops laundered money from going through protocols. It stops hacks from happening before they hack.

Well, before the users have hacked. So what we're going to talk about today is the LP exit because this the signals that happen that traders look for they're actually happening before the actual trade has happened on chain right we all have heard of Bloomberg terminal and there's a reason for this there's a reason that these products can sell and the same data lives on chain the market is coming there this a little bit what I'm talking about right I've gone through the three models in the last slide Um, and this is what blockite does. This is how we're using AI on chain. Are this is where the value comes from from our products. I'm going to skip some of these.

So, how did this idea come to life? This is one of my co-founders. Um, he is the uh our chief scientist, but he also has been um a lot of his work is published on the Harvard data versse. And this is because a program called SWAN SF is has been used by the states to essentially save astronauts when they leave the Earth's magnetosphere. Right?

It's very dangerous if a solar particle happen or a solar flare happens to occur. The solar power particles, they can kill you. So you only know in a couple seconds. It travels extremely fast. So what they've done is they've looked at regions of the sun.

They've um they've used CNN's so neural networks to analyze exactly, you know, how these events have happened over a 10-year span and when they're going to happen next. This is what we've done with a graph neural network on chain. So these are this is a little bit more about Swan SF. So there there was a lot of data as I said occurring over the um they collected data over 10 years and they they broke it down into 12minute samples. So the same architecture is essentially what we use to build our AI.

This is not a wrapper around chat GPT or anything like that. So as you can see from this slide, this is just a little bit more information about what we're taking in, right? So we have transaction frequency, token flow, um, governance, cross out. We're looking at the block level. So we need to understand when things happen, especially in cases of fraud instantly.

We need to understand where the source of funds are coming from. So is this a new wallet? Has this wallet, who is this wallet interacting with? Does this wallet want to move a large or small amount of money? And all of these factors are things that weigh in on whether or not a protocol is interested in accepting this transaction.

And the protocols that work with us, they can adjust their risk rate based on all these factors. It's a very modular system. So we're looking at a few questions, right? Is this going towards churn growth or anomaly in terms of the churn um model? But in terms of our LP models, we're looking at their behavior and what they've done in the past based on both the sentiment of the market and in general the fluctuation price of the token or the types of tokens that they are taking in and holding or dumping.

So, we've used a lot of uh different models on this. Um, as I mentioned earlier, we're using a graph neural network um because it's the only one that really makes sense here. A CNN is mainly done for uh photos. So, whilst you can take photos of the sun, you're taking photos of the blockchain is kind of ridiculous. Um, support vector machines.

So, we're segmenting the wallets. Um, this is used for a lot of different things, right? So, we want to know if this is a bot. So, you can can pretty much tell from the behavior, right? If this is a bot or a human and whether it's a safe wallet or at risk um or at risk of being fraudulent given the wallets it's interacted with.

The decision tree is a lot of what we're using too because they're sort of like if then rules, right? So, we want to know if this is a new wallet, what it's holding, ho how it's communicating with other wallets on chain and multi-layer um perception. So, this is a like nonlinear um compound patterns that look at all of the features at once. That's another neural network. So, Why I'm talking a little bit about the graph neural network throughout this talk is because it's it's different um from the way the sun works in a lot of ways because their main we we base this work off of swan SF right so they're the what the sun is doing is mainly um like the how we're when you would use a graph neural network for that it's more like how the magnetic field reasons don't just have properties but they they connect to each other and they have these relationships to each other.

But on chain, so let's let's talk about farming. So one wallet, if you're just looking at a singular wallet interacting with a protocol, it might be making small transactions here or there or voting occasionally, but when you have when you look at a bunch of them, maybe 50 together, you can see through these patterns that they might be farming this protocol. So, I'm going to skip a couple of these because my designer had a death in the family this week. So, some of these aren't on par with really what we're talking about today. um because we're mainly talking about the uh the trading, right, and how we see things before they happen.

Um and we're looking at that in relation to our uh our AI work and how this has been adjusted over um from SwanSF into block site. So when we look at the original work that came from Swanf, we see that there are 51 parameters. But when we look at a wallet and its trading, we're looking at at um these transactions and how they're happening on the blockchain. We're looking at um the how they've reacted in a time series to this particular wallet, the onchain behavior, how they've reacted over the years to different events, right? So we can use the Twitter API.

We can use a bunch of APIs to then pull in this data about what's going on in the market offchain because those are the two factors that are really going to influence traders and wallets and how they behave. Human behavior. So [snorts] what we make decisions by and it's likely that most firms most liquidity providers will follow the same behavior patterns. And we see a lot of the times too once the incentives are pulled away. So for example like uh we have a market maker an LP that is on a token loan deal and they know another one's coming.

So they may not dump all of their tokens all of the project's tokens until the next wave of funds comes in. then they get another batch of those tokens and now they're dumping them at this level. It's just not acceptable. Um, this also happens with trading because you know the traders want insights. They want to see things before they happen and you and me sitting at our computer, we don't really have the same amount of tools that these larger firms have.

So, we're at a disadvantage. in terms of long-term trading um trajectories and understanding our portfolios. And that's not necessarily how the market should be split. We should actually have the rights to know what the traders know, look at what the traders are looking at and understand and identify the signals to the same degree. So what we're trying to do is not replace entire teams of quants, but even the playing field with predictive modeling.

And I apologize for that. My computer doesn't listen, but it's coming back on. So, I could talk to you all day about different parameters that we've used, but a lot of you I feel like are crypto native and understand that. But let's look at um one thing I haven't touched on yet is an XAI layer. So we're looking at essentially what that is is it's has nothing to do with Twitter but these are explainable AI patterns.

So AI doesn't become a black box but it becomes um we sort of break it down into the math and in a way that can make this really easy to understand is we want to know at which point the trajectory began and that helps projects understand why. So if we know that on September 23rd, throwing out a random date as an example here, a bunch of wallets churned all at once, which in turn churn in this case means that your users are no longer retained, which user acquisition across the board in tech is incredibly expensive. So if you are losing x% of your users in in bulk at any given time, this isn't good. So this is just an example, right? But let's say you have, you know, a large churn on a certain date.

You can go back and you can look at the sentiment. You can look at what your company announced that day. And this is this kind of information. It helps the small guy. This is why I got into crypto because it was sort of a a door opener for anyone who wanted to maybe have an IPO or get into business or even just trade fairly and have an open market.

So, I don't understand why a small company shouldn't have all the resources that a large one does to to have user retention because if you don't have any users, you don't have a product. you don't have a company really. You don't have any sort of income. This is important. And also on the user side, they deserve to get some something that they want for their product.

All right? So, we're looking at patterns, right? We're look, we read what the onchain data already shows, but we one thing I want to emphasize since we're at a privacy conference is we are not um singling out single people, right? Unless they ask to be. But this is a totally different thing.

If you want to prove to a bank that your source of funds are good, great. Come to us, talk to us. But we're looking at behavior for we're looking at behavior for companies. We're looking at behavior for anyone launching a token. We're not looking at single pe single individuals.

We've had VCs ask us if we will take in the data from web browsers. No, we won't. That's unethical. That's where the line is drawn. And ultimately, people should be held accountable for their behavior.

I don't really think that's a debate that we can have. If you want to have it with me, we can after I'm done talking. But yeah, so this is just a pro. It's like a line of defense. It's a protective line of defense is what blockite is for protocols for people who want to invest in something for banks so they don't lose their license for small banks so they don't lose their net license because they're not all JP Morgan and they can't all afford a team of developers to come in and help but yet they still want to give maybe lower interest rates or for somebody looking to buy a house who wants to take their uh down payment off chain.

These things can happen. We just need to make sure the tools and resources are available. So let's talk a little bit about the explainable AI. So we have we have SHAP um and that is looking at the distri uh the dis how um how predictions credit across different features. So the weights of each feature.

So for an example for churn, right? So let's say we are looking at a group of users and they stopped engaging with a DAO at a certain point. Um and we say okay this is the there all of your users the biggest signal for their disengagement is the is the Dow use. So it's quantifying what behaviors are causing this and this um grand cam and it it highlights which regions of input the model focuses on. So it shows clusters and what what is exactly driving the behavior um and integrated uh gradients.

This is the baseline path I was talking about with the September 23rd example. Um, so let me just skip. Sorry guys, I'm using a let me just skip to the end here because I'm at time. But the point is of all the of the discussion I'm having is that at this point, you know, that privacy kind of connects with how we connect with our users at the same time. We can't necessarily say okay all of your transactions in any monetary form whether that's banks etc.

They they can't hide the transaction. Banks can't you can't either on chain. There is a point when regulators will come in and there is a point that you need to take some responsibility but we can hide to a degree a user's privacy without naming them and still give the companies the access they need in order to be successful. and and that's where the moral ground lies here for block site. Um, and this also is, as I said before, what we're looking at at scale for traders.

So, we're looking at as a whole how behaviors are impacting each asset in a portfolio.

[applause]