Last week, Nuklai CEO Matthijs de Vries joined leaders from Botify, Nexera, Inflectiv, and Vanar in a roundtable discussion hosted by GraphAI on AI, RWAs, and the future of onchain intelligence.

The session explored how projects across the ecosystem are tackling the challenges of making AI more reliable, scalable, and useful in real-world financial and enterprise contexts. Matthijs spoke about Nexus, Nuklai’s AI-native query engine, and how it integrates with GraphAI’s structured subgraphs to create a unified environment for AI agents to reason over both on-chain and off-chain data.

Listen to the whole recording on X. In this article, we’ll highlight key discussion points, and an abridged transcript of Matthijs’ responses in the AMA.

Key Insights

• Making data approachable: Nexus is designed to make it easier for both technical and non-technical users to query diverse datasets in one place, starting with structured data.
  
  

• Metadata over raw data: By training LLMs on metadata instead of raw data, Nexus avoids hallucinations, scales across thousands of data sources, and ensures enterprise-grade data security.
  
  

• Onchain × offchain synergy: GraphAI provides structured onchain intelligence, while Nexus connects enterprise offchain datasets — together enabling a “single source of truth” for use cases like compliance, fraud detection, and financial audits.
  
  

• Verifiability as core: Using Helix, Nuklai’s L1, queries and results can be hashed and stored on-chain, ensuring that AI decisions are transparent, reproducible, and trusted.
  
  

• RWAs and adoption: Transparency is the missing piece for RWA adoption. By letting users interact with complex asset data through natural language, AI can lower barriers and expand access to tokenized assets.

To make the discussion easy to digest, we’ve included an abridged transcript focused on Matthijs’ responses below, with the full recording linked for anyone who wants the complete session. Huge thanks to GraphAI for hosting, and to Botify, Nexera, Inflectiv, and Vanar for a sharp, forward-looking conversation on the future of AI, RWAs, and verifiable onchain/offchain intelligence.

Abridged Transcript

Introduction to Nuklai

I think first and foremost, when we speak about Nuklai, it's about making working with data easier. Our end goal is that the world's data — and I mean that with all the business data that's siloed right now or maybe your own data that you generate with your car, data like that; and also reference data that just is country codes, colors of flags — available as a single piece ontology and the world can talk with this data and leverage it for whatever use case integration, a research paper, whatever. That's the end vision. And we do that step by step. And the first step is Nexus.

“When we speak about Nuklai, it’s about making working with data easier. The end goal is that the world’s data — from siloed enterprise data to simple reference data like country codes or colors of flags — becomes available in a unified ontology that anyone can leverage. That’s the vision. The first step is Nexus.”

Let's take a step back. When we speak about data it is in different forms and shapes. It's stored in different sources. So when we want to work with the world's data, we need to make working with the data as easy as possible.

Now, we focus first on structured data. Structured data is what we see typically in spreadsheets. It's predictable. You have one row of data that's in a specific format. The next row is different data, but it's in the same format. If you have one column that says it's an amount in dollars, it's not going to be information about the tree on the next record. You know, it's going to be another amount in dollars. Different amount, but you know, an amount. So that's structured data.

And structured data is typically in databases, relational or NoSQL, or it can be in files like CSV files, spreadsheets. Working or getting access to the data, tapping into it, typically requires integrations that require software developers, require technical knowledge, resources, time, money. For one source, it's doable. You have a lot of free software and stuff to work with. But if you have many different sources, or even a few different sources, and you want to talk to these different sources at once, it is going to be very difficult. There's no way, no protocol, you know, besides SQL, which doesn't support all the kinds of data sources, to tap into this data at once.

So we created something unique in Nexus, which allows you to query all these different kinds of data sources as if they are the same type of data in the same data source in one single way. And then what we do next, so this allows you to talk to this data in a technical way, you know, as a developer, super easily, but it's not good enough, right?

So if we want to allow people that don't have any technical knowledge to talk to this data in as easy a way as possible, meaning natural language, we need to introduce generative AI. LLMs in general are very powerful in order to translate natural language into something technical.

So that's the first step. That's already done.

“Structured data is predictable, like spreadsheets or databases. But getting access to multiple sources typically requires developers, integrations, and time. Nexus changes that by letting you query across many sources as if they were all one — and in natural language, not just SQL.”

Then what we typically see with competitors, alternative solutions, is that they take the data itself. And train the LLM on it in order for the LLM to understand. And we say that's not good enough, because if you have big amounts of data, it will only lead to hallucinations. And hallucinations is a very important topic right now, because if we see LLMs, we don't always know if they gave the right answer, right? Because they hallucinate a lot.

So we chose to, instead of training the LLM on the data itself, we train it on the metadata.This solves two things. The LLM is not aware of the data itself, so that keeps the data safer, which is very important for enterprises, governments, healthcare, stuff like that. And secondly, it makes it extremely scalable. You can literally add thousands of data sources at once. And just because you scale on the metadata, the LLM will always have the right data, in real time without having to resync the RAG, the data that's being fed to the LLM.

So, we created that AI layer, that query layer, so that people without any technical understanding can start talking to data sources using natural language on an extreme scale. So yeah, that's what we do with Nexus. And when it comes to the layer one infrastructure is when it comes to when the AI has to make decisions based on what specific data is like a number or an outcome from a calculation, then you want to store those values somewhere in an immutable ledger to be 100% sure that the decision that the AI made was based on something factual. It wasn't made up.

So that's where decentralization really comes into play in an actually really powerful use case.

“Competitors often train LLMs on the data itself. We chose to train on metadata. This keeps data safe — critical for enterprises and governments — and makes the system scalable. You can add thousands of data sources at once, and because it scales on metadata, answers stay real-time without resyncing. It also avoids hallucinations, which is a huge issue in AI today.”

How does Nuklai’s off-chain datasets combine with GraphAI’s on-chain intelligence, and what does that mean in practice?

Nuklai is data-agnostic. We are LLM-agnostic, we're data-agnostic, so we don't really, we don't define what kind of data needs to be on board for what kind of industries. We provide merely the infrastructure to it and the technologies to give utility to that data.

Now, when it comes to the data sets that we typically onboard, it's about business data. And we see a lot of use cases around on-chain data. And we don't necessarily want to reinvent the wheel and put on-chain data on Nexus.

“We’re data-agnostic. We don’t reinvent the wheel for onchain data — GraphAI is built for that. By integrating, users can query both onchain and offchain data together. Think DeFi, RWAs, or financial forensics: combining datasets into one truth.”

So an integration with GraphAI is around the fact that people can easily create graphs or subgraphs in order to query on-chain data for specific use cases. I saw mentions of DeFi and RWAs, and in our case, with the businesses that we speak to, it should be more around forensics, fraud research, you know, the more things for financial institution enterprises in order to combine off-chain data that they've collected with the on-chain data using GraphAI through Nexus. Let's combine these two data sources as one single truth.

What do you see as kind of the most exciting real-world asset or I suppose enterprise more generally use cases for this kind of intelligence technology?

Yeah, there are a couple of things. I just recently, last week, had a call with an enterprise around, it's a big consultancy firm, they have a lot of financial clients, and they were mentioning that one of the biggest problems their clients are facing is when they have audits, you know, all financial companies get audits, right? And during this audit, for example, if a certain number shows up, the auditor would ask “How did you come to these numbers?” Let's say the number says 3 billion. Where does this 3 billion come from?

Well, they could say, you know, we use Nexus to create a query. This query took data from this data source, from that data source. And while querying it, you know, these amounts, you know, they came out of the data sources. And applying this query, there was a calculation and that resulted in 3 billion. That all sounds cool, you know, unbelievable, but... How do you know at the time of the execution of the query that data was really that data?

So that's where decentralization comes into play, right? So for enterprises, a real-world data, use case is having that snapshot of the data hashed, so it doesn't need to be literally the data itself, combined with a hash of the query or the algorithm, combined with a hash of the result on-chain, would be proof to the auditor that data was really the data and that $3 billion hasn't been manipulated sometime after.

“When AI makes decisions based on specific numbers or calculations, you want to be 100% sure those values are real and immutable. That’s where decentralization comes in. With Helix, we can store hashes of queries, data snapshots, and results on-chain as proof. That gives enterprises confidence that the $3B number in an audit, for example, hasn’t been manipulated.”

How important is verifiability or transparency for these AI driven workflows? 

It's the most important. This is the biggest hurdle of generative AI right now, right? We see it in image generation. It makes stuff up. It makes decisions beyond our control or counter to what we instruct the generative AI with. And when it comes with answers, sometimes they're completely based out of nothing. And when you call out the LLM on it, they are like, “Yeah, oops, that's on me. I thought just, you know, it's a nice number”.

We need full verifiability and traceability down to the source of the data, in my opinion, in order to be able to be confident. In the accuracy of the generative AI outputs. So when it comes to using generative AI for complex reporting, business intelligence insights, or even decision making, you know, when the LLM using MCP (Model Context Protocol) Tools where it can actually execute actions, we need to be a 100% sure that the actions that it took were based on actual factual information, factual data from trustworthy sources.

“Real-world assets are centered on data. The value depends on the quality and transparency of that data. Generative AI can make RWAs more approachable by letting investors or enterprises have a natural conversation with the data — breaking down complexity and supporting broader adoption.”

So Nexus, you know, when speaking about Nexus, it's traceability at its core. Every answer an LLM gives can be traced back to its source and can be replicated. So if it creates a query, if it executed that and data came out of it, you can replay the query on the data source and see if, you know, that matches what the LLM the output was.

Having, you know, in some use cases where it's very important to, because it's mission critical, you would ideally need to have those kinds of steps, the inputs and outputs of these tool callings on-chain, even just as hashes if you don't want to show the actual data.

Verifiability and traceability should be at the core of mission-critical AI solutions. And that's exactly what we decided to do. And that's how Helix, our layer one network, is complementary to Nexus.

“Traceability is the core of Nexus. Every answer can be traced back to the source and replayed. For mission-critical workflows, inputs and outputs can even be hashed on-chain. Verifiability and traceability must be at the heart of any serious AI system.”

What do you think is the most important problem that still needs solving before the space can really scale?

For us it’s very easy. RWAs are centered around data. The data determines the value. Of the asset. And there can be many types of data. So when I see, you know, in many cases where lies the challenge for the end user, the investor, whatever. It's around transparency and understanding of what that asset is about.

So having AI, in our case specifically, that has access to all the data surrounding the RWA, and being able to just have a conversation around it, you know, have a better understanding, especially if the AI understands the user querying the data, for example, risk profiles, kind of, you know, what kind of, you know, what kind of person it is. It could give advice on what kind of RWAs are more fitting to this person, how to approach it, what's risk-reward, stuff like that.

So for us, it's around that data. Make it approachable instead of everything that's really opaque right now and seems very complicated to people that are maybe new to Web3. And if we want to speak about the adoption of Web3 and adoption of tokenization of RWAs, we need to make it low barrier and as easy to understand, and I think generative AI is going to play a very large role in that.