NODES AI 2026 - The Future of Enterprise AI Is Context Infrastructure, Not Better Models

This talk covers the importance of using context infrastructure for enhancing AI agents in organizational settings. The speaker, Manuel, an AI and product specialist, explains how businesses often suffer from fragmented data definitions and siloed AI implementations, which leads to underperformance and confusion in AI outputs. He introduces Spintel as a solution that allows for a centralized graph database to integrate and maintain business logic, making AI context portable, vendor-independent, and governed. By employing ontology-based structures and temporal memory, Spintel empowers organizations to build custom AI assistants capable of streamlining operations across various use cases, including content creation and customer support. The session includes a practical demonstration showing how Spintel facilitates the creation of AI modules tailored for different business needs while ensuring the relevancy and quality of the generated content.

Thanks, Chelsea. Um, well, first of all, thank you all for being here. and um, yeah, this session is going to be all about uh, contest infrastructure. And um, and how um, how to use it to enable uh, AI agents in the performance of their task, no? before getting started, just a bit of an introduction about myself. Uh, well, as you see, my name is Manuel. Um, uh,

I am an AI and product specialist, uh, mostly working as a freelancer and consultant in um, helping organizations to uh, discover, build, and deploy uh, AI systems for internal or external use. And uh, my experience with graphs, well, it all started in Clarida. Um, with the with the um, uh, build-up of the of the uh, centralized uh, uh, source for for for the for the entire company.

Um, I've also been working in in in um, uh, graph power systems for SMBs uh, to to um, streamline operations. Um, and of course, Pinterest, which is the the uh, the topic of this uh, of this session. Um, apart from that, I've also been working in in in in uh, uh, AI-related stuff within like lastly in in in marketing. which is definitely a I an industry that

can benefit for for obviously for for knowledge graphs, no? first, let's address the the the problem space. So, as you all know, um, many of the people here know that contest fragmenta- fragmentation is is uh, is one of the of the problems uh, that graphs can solve. Uh, because if you if you are in a big corporation, you would recognize that uh, definitions and categories do not

happen to be centralized, no? Um there are in some teams they call, for example, customers is something and in another team's customers is another thing. That's just a simple example. Um and the say the tries to get all that data out of the box, all documentation with different categorization, etc., into an AI that would produce um underperformance, no hallucinations, because the AI would be confused, no, in

that in that sense. Um and uh so that is that is the the the the first problem. And then the second problem that we see is the the problem the problem of the silo, no, the silo problem. this we see a lot with with vendor-provided AIs, for example, for CRM systems, SEO systems, uh etc., etc., where the AI implemented is uh kind of like trapped in in

a specific business logic, no, it's it's kind of like set up exclusively for that. Um and you cannot move it between different the CRMs, no, because it's it's it's siloed there. Uh and this also applies to to custom AI assistants, no, for example, in in uh if you if your company you happen to have like ChatGPT Enterprise or or or Cloud Code or or or Gemini, um

then each of the custom assistants created are siloed are kind of like uh blocked within within the the the the context of the snapshots that have been uploaded into that uh specific assistant, no? Um and that is that is that is a silo, of course. having like knowing this this context, there is of course we live in a now in a in a paradigm shift that everyone

recognize, no, that is that models are replaceable, so LLMs are replaceable. But the business context is not. Uh Uh, business context always stay the same and and and evolves, of course. And so, yes, AI becomes commoditized, the real enduring asset for any company is their their uh lies not in the LLM itself, that they whether they use the most expensive model or not, uh or in the

prompts, but rather in the in the evolving context of that company, So, the solution that I've been working with with uh with Spintel is a uh so, keeps the business logic portable, um governed and vendor independent in a sense that uh you can like export it and also plug any any any um different LLM you have in order to perform uh the task that you want to

perform, no? And that's the solution that that uh that uh Spintel is. Um as you can see here in the graph, uh like in you will have like of course in a in a in a given company, no? You have like your your your your current tech stack with like with the documentation that you store it on in Google Drive or in Microsoft Office, Notion, emails, Slack,

And then uh via integration or via manual upload, then it would go to the to the uh single graph of the of the company, this enabled by by Spintel. And then via via that users would be able to create specific, let's say like modules, custom AI assistants, that would not require you to worry about um let's say uploading the necessary context because the context would be like

already like maintained in the in the in the in the graph, and you just have to like pass like minimal instructions and the data points in order to to uh output whatever you want to to to output, no? Now, the four pillars from where this is this has been built is pretty much like uh so, first it's ontology-based. Every time a uh it has As integration or

a document is uploaded, it will first generate a schema and ontology that AI would do Um and then after that it will it will populate the the with the actual data. It has temporal memory um as uh Lasse Andresen mentioned in the in the in the talk in the in the in the beginning of the session. it gives like timestamps to every time something is uploaded so

then AI knows what has been the temporal in the evolution and what is uh updated and what is for example like really old. Uh governance execution in a sense that is not only um it doesn't enable enable only read but also enable to create uh update or or or delete data in the in the graph database. Um and it's portable because uh in a sense you can

uh export all the data always uh in a standardized format so you can use it for for any uh other external system or or any other purpose that you might have. Um so that is that is after my presentation without further ado I will jump to the main um to Spinte the tool. Um as you can see well it has been built as really simple. You have

like the the the data source section um where um you can upload the the necessary documentation that uh or the critical context that you currently have. In this case I already have six documents that have been that created with with dummy data. Um which is about Lumio is a is a is a dummy company. and then what it happens here is that at first as I mentioned

uh it's ontology based so in a sense at first it will go through an AI data modeler, uh, agent which, uh, will, uh, map out what is the schema or the the the the ontology of of, uh, of, of the context contained into those documents. It will then, once that that is finalized, it will jump towards the to to generating the data on the back of the

schema or the ontology that's been generated by the previous agent. Um, and, um, and then it will populate the the the graph. the schema and ontology is is is stored in the same graph. So, that means that it's not stored in a in a in a in a weird back end or something like that. It's like directly stored in the graph. Um, and and you can see

here, uh, how all the graph visualization, um, this is the meta schema. You can see here like everything that is as it goes. Uh, by the way, the data visualization I used, uh, Cytoscape. Um, they're open source and and they have a really really, uh, a huge variety of of, uh, visualizations, not only for graphs. So, if if, uh, if any of you is interested in in,

uh, in, uh, implementing it, check it out. That's really worth worth, uh, worth checking. Um, and I think that this is now completed. You can see also it's been done the chunking, etc., etc. you can also see like here in the in the window, um, like all the the entities that have been created. Of course, the the the product resume, you can see the also in the

description, um, the properties, um, and etc., etc. and then you meet, of course, you can see this the schema. You can see like all the, you know, properties and and so forth. So, now, once the once the the the the the central source is already, has been created, then we can proceed to to, uh, query the graph or create the different different modules in order to perform

specific tasks, no? in this case, I'm going to just confirm that everything is uh well reachable by the AI, so I'll just like ask a test question, what is Lumio tone of voice and principles? Because of course like what I have I've uploaded is marketing marketing material for for this company. Um so then what I'm asking here, let me get it ready, is is um what is

the tone of voice to see what it extracts and how it extracts it, Um you can see this is uh is reasoning. Um it's calling I research and also text to cipher. With the retrieval, how it works is that the first uh you see the combination of of course I research I research which is vector index and and full text search. Um and also uh text to

cipher in order to traverse through the through the um entities that has uh that has been fetched by one retrieval or the other. and we can see how is it going then it effectively um gets the uh tone of voice for this company. Okay, then we have also here like what what tools has used and and of course also the citations not like here not the citations

because actually also returns the the source of all the um uh response that has been given. so in this case you see that this is this is just like the the the default out of the box uh uh chat. Um and now what I'm going to do is I'm going to create modules for for this. Um a module is a custom AI assistant in this case. Uh

so what I'm going to do for example I'm going to create a content creation assistant. If if any of you watching this have experience in marketing, this is This is uh for for Yeah, for creating on-brand content, which which is kind of like creating um let's say uh product listing pages or or uh blogs which align with the with the uh with the tone of voice of

the of the of a given company. For instructions, I'm going to paste also this uh these instructions. We say like you are uh Lumio's content creation You write blog posts, LinkedIn copies, uh email sequences, and landing page content. Uh I will follow always follow Lumio's tone of voice guidelines, use uh use only approved uh call to actions from the call to action library, write for the persona

specified in each request, back every every claim with a valid validated proof point from the product documentation, um never use words from the avoid list. And then something something to also uh make it more efficient um is that I also give the possibility to point to a specific entity or endpoint from which uh this module will always have to to um to return kind of like to

fetch first and then it traverse through that central entity. Um in this case I have here like uh from initially in the data source, I can select here what is the the the central uh let's say uh entity. And in this case I'm going to take content page, for example. And the recency, because of the temporality, I will take uh for example the last 30 days. So

then I created a I created the module, is this one. Um and then what I'm what I'm going to ask it to write a LinkedIn post for Lumio. So, this one. I'll say, "Write a LinkedIn post for Lumio targeting people analytics leads." Uh this is one of the personas that we uploaded uh in the uh in the data source. And then a hook first under 100 150

words. So, there it goes. Processing the the request. Um it's going to pull everything. It's just over here the reasoning in the activity. It does a lot of a lot of calls. It could It might be of course this There's still a lot of work to do it to make it more efficient in the calls. But in any case, the most important is that it will it

will it will work, no? So, that's that's uh let's see. What it does. We would There is a problem that I'm having. I am here. Um so, then you write the the LinkedIn post for James O- Okonkwo, which is uh correct. And it creates uh yeah, like a LinkedIn post. And there's also editorial notes and and and so forth. This is This is really valuable because in

my my previous experience also within within the the the marketing industry, for example, um you see how companies uh or agencies that offer AI solutions, normally they um uh they create a kind of like AI tool where all the instructions, all the context is is kind of like done in the back end. Um and then only the specific context for that like or for SEO is is

added, no? But, there is missing a lot of times also the con- the context from other parts of the documentation that are also relevant for this. Like, for example, for some information that you need for CRM is also uh relevant for for for creation of content content, no? So, then this this helps a lot. This provides it like as much more context than than uh to to

make uh the quality better. and and another benefit is, of course, something I've seen in within uh the marketing industry, is that there's different tools created for different things. So, you have like one tool for SEO, one tool for CRM, one tool for other things. And then they're three like separated systems in order to to do something that can be unified into one single uh uh let's

say uh central context tool that where you only need to upload the necessary uh uh context, and then it just creates the necessary modules in order to perform a task, no? And that's something I'm going to show now. So, we created a a kind of like a marketing content crea- content creation tool. Um uh sorry, module. But then the next uh without having to create a new

uh different tool for, for example, other things, you can easily just create a a another module for, let's say, um uh customer support assistant, no? not something practical to do to do within a company to create a customer support assistant for internal use, but this is just to show that it's possible to also just create a a module that can be used externally, you know, by by

clients. Um so, that is that is what I'm going to do. I'm going to create a uh customer support assistant. Um for external use. The the content creation was for internal, this was for external. Um and then I'm going to copy the instructions, which I already had saved. Uh which is a YourLumi customer support assistant. You answer customer and partner questions uh accurately and concisely. I will

answer the questions first before adding context. Be specific, use real number, real timelines, real feature names. If a feature does not exist, say so clearly and offer a close the closest available alternative. Never speculate on road maps. Never never promise pricing or discounts not documented in our space. Then I will just um add here as a topic seed for example which one can I use? But I

will I will use user because user of course is the the the um no, not user. I will use a product. Okay. Um this is like a customer support is like obviously about about the product so far of Lumio. So then I create the customer support assistant and then I can just like replicate what what uh uh a customer external user can can for example uh ask,

you know? which for example this one. Um we are on the starter plan and want to access Lumio's narrative. What are our options? This could be more more like a partner question. but it it's also the the the purpose. Um so then Lumio's narrative is not included on the starter plan. Based on the available data, narrative is available in connect connect and inside plan. Um to get

access your option to upgrade to connect by which is uh correct. Um and also provides next steps to to for the customer to see see pricing and and talk to the team. Um it does use only average search and and you can see also the um the citations of where where citation the the sources from where it got the all this information, you know? that is that

is that is kind of like a a a quick overview of of what what you can and in this case I'm going to go back to So that has been the the quick demo. Um Thank you for watching and well I guess because I have I still have time. I guess I can answer also the questions. What is the link to get more info on Spintel? Um

yeah I like you can that that is the the correct link but if you want more information you can contact me in LinkedIn. Personally I will actually share my LinkedIn URL in the in the comments here so you can have it. Okay so that is that. Let me see. Yeah I think that's that's all. There's no more questions.