NODES AI 2026 - Aura Agents in Action

This talk covers the capabilities of Neo4j's Aura Agent, which enables users to build and deploy AI agents on top of knowledge graphs quickly and easily. The speaker, Ed Sandoval, explains how users can create these agents using the Aura console, incorporating tools for parameterized queries, similarity searches, and natural language processing to enhance user interactions. Following this introduction, representatives from Napsack Foundation and Lean Mode AI share their experiences and applications of Aura Agent in public safety and software engineering, respectively. Napsack Foundation discusses their focus on using geospatial tools to support emergency response and local infrastructure awareness, while Lean Mode AI reveals their approach to improving software development workflows by analyzing bottlenecks and enhancing decision-making through data integration. The session concludes with details about an upcoming Aura Agent Hackathon encouraging participation and innovation among developers.

[music] >> Welcome everyone and thank you so much for joining us today. My name is Ed Sandoval and I'm the product manager for Aura Agent. I'm super excited to be here for this session, but what excites me the most is not what I have to say, but is what the folks from Napsack Foundation and Lead Mode have to say. We have a lot for you today. I'll

kick things things off with a quick overview Neo4j Aura Agent, then I'll hand it over to Kevin K and Jesse Osborne from the Napsack Foundation. They're a non-profit that provides geospatial tools to public safety organizations across the USA. Then we'll hear from Diego Giron, founder and CEO of Ling Mode AI. Diego and his team are building a product that allows software engineering organizations to connect tools like

Jira, GitHub, Slack into a knowledge graph. They use Aura Agent to power AI assistants to their user. And at the end we'll finish the session with details of an exciting competition that we have for you. So, to set the context and for those of you not familiar with Aura Agent, Aura Agent is a new product available in the Aura console that lets you build and deploy your

AI agents on top of a knowledge graph in minutes. You can create the agent, add tools and integrate it into your application with simplicity. Uh in order to bring your Aura Agent to life, you start from a knowledge graph. And this knowledge graph has to be in AuraDB. It doesn't matter what kind of graph do you have, whether it's a context graph, memory graph, domain graph, a

graph you build with from unstructured data sources, a graph built from data in relational databases, it doesn't matter. As long as in AuraDB, you're good to go. The second step is that in the Aura console you can create the agent, set it up with tools, and test it. We all know that an agent is only as good as as the tools available to it. And for Aura

agents, we provide a specialist tools to retrieve data from your knowledge graph. Your agent can have three types of tool. The first one is parameterized query templates or Cypher templates. You can use these types of tools when you can anticipate the questions that your users will ask. >> [clears throat] >> Something like, "Give me details for this patient." or "Get me the details from this contract." The

Cypher template will help your agent respond reliably pulling precise information from the Um the second type of tool is similarity search. This is your vector search tool. The user asks a question, the system find the most relevant entry points in the graph, and your agent explores nearby nodes and relationship to build better context and hopefully richer answer. The third type of tool is a text to Cypher

tool. You can use it for ad hoc questions that you're you cannot fully anticipate. This tool the agent takes the natural language a fine-tuned model, and the schema of your knowledge graph, and execute a generated query, Cypher query, run it against the knowledge graph, and formulates an answer. that's it. Once you've done adding tools to your your agent, the next step is evaluation. Once again, in the

Aura console, you can fire questions to the agents, check the response set, check the reasoning, verify all the tools that were invoked in order to answer every question. This is an iterative process. eventually, once you're happy with your agent, you can connect it to your application in two main ways. Uh you can use every Aura agent can be exposed as a Aura API REST endpoint. And with

that, you simply get your app to interact with the agent via the REST API. And every agent also can be exposed as an MCP server endpoint, which means that you can use your MCP client of choice, Cloud in this example, to use that MCP server endpoint to use the agent in your Cloud work. And that's a quick overview of Aura agent. But now, I'd like to hand

it over to people who are actually building their products using Aura agent. So, please welcome to stage Kevin Kay from the NAPSG Foundation. Kevin. >> Thanks, Ed. Uh appreciate it. Uh I'm Kevin Kay. I'm the Chief Operating Officer here at the National Alliance for Public Safety GIS Foundation. Uh I've been with local government and public safety for the past 15 years, and in the last 10 have

really tried to explore local infrastructure and the connections amongst them. We do want to start out with a big thank you first to Ed and the Neo4j team for having us on, but more importantly to the public safety community who is our main community that is using our technology. Uh for those of you in and outside the US, this includes law enforcement, fire, uh urban rescue, EMS,

911, emergency management, and they use technology and data every day from small incidents to large disasters. We've made a lot of progress in the geospatial space, but we see a new avenue of using graph database and graph data science to really improve their response to incidents and help the outcomes for both the responders and survivors. So, again, that's our community we're building for, and thank you again

for having us. Next slide. So, a bit about Napsik Foundation. Founded over 20 years ago, we are a 501c3 nonprofit. Uh our main goal is to empower public safety personnel with geospatial tools and other advanced technologies to better prepare for and respond to events. This spans multiple disciplines, spans multiple categories of building tools, training and exercise, and technical support. But more importantly, we through the generous uh

funds of both our sponsors and contracts we have with federal organizations provide it all at no cost. So, we're not selling anything, and we're a true neutral party when it comes to an R&D space for public safety. So, I just want to talk a bit about the gap and how we ended up here and talking about our local infrastructure network knowledge tool. We have multiple projects and

multiple incidents that all uncover the same challenge. That local jurisdictions struggle to identify infrastructure, how they're connected, and cascading impacts. So, whether it's from ongoing projects, world events where you see attacks on infrastructure, or even recent events here in the United States such as the LA wildfires and Hurricane Helene, almost every incident has some sort of cascading impact related to utilities. And these are lifelines, water, wastewater,

transportation, energy, hospitals, government. The struggle here is that in the United States, the majority of these are owned by private sector entities. Uh so, it's kind of difficult to understand what exists and how they're connected. So, there's definitely existing tools at the federal and sometimes state level. Uh these tools prioritize nationally significant infrastructure. It's typically secured at a higher level. It's used for large incidents and used

by those with vetted access. That is not what the locals are looking for. Uh we've heard a lot of feedback that the locals want tools to at a base level identify what exists in their jurisdiction by certain sector and by a certain taxonomy and then how they are connected. Then they could conduct outreach. They can do physical security assessments, cybersecurity assessments, bring them into their emergency operations

center and do exercises see how they would react to certain incidents, but that does not currently exist at the local level. So, again, that's the origins of LINK. Uh Jesse's going to walk through a bit of the technical side, but at the end of the day, we really want something that locals can interact with and make sense of their data in a easy-to-use way. Agents have allowed

us to in a plain language way explore the complex nature of interconnected infrastructure. So, with that I'm going to hand it over to Jesse to talk about the technical pieces. >> Great. Thanks, uh Thanks, Kevin. Um actually, can you hold off real quick? Uh Ed, I'm going to give myself a quick introduction. Hi, uh my name is Jesse Osborne. I'm a data scientist and uh the lead

architect for the LINK project or local infrastructure network knowledge uh web application that we've been developing uh that Kevin had uh gave a good introduction to. Um I think the best way of being able to um get to know and understand link is is through seeing it in action. So, I put together here a quick demo. You can go ahead and start the demo now, Ed. Of

just how easy it is to use link. Simply just logging in and you're immediately able to take action and ask questions in natural language. So, link uses our agents and an RDB instance under the hood. In this case, the user has asked the question of can you show me all of the power substations? the DB behind this has just a single county within Colorado. And as you

can see over on the left-hand side, a number of key properties of the power substations over on the bottom right. A graph and a synchronized map to show the geographic context. User is able to follow up with some follow-up questions. You know, which one of these are connected to hospitals? Now, you're seeing some some real graph in play. With these relationships, some of them being physical relationships.

You can then interrogate the data and learn a little bit more about all the different elements that you're seeing within the the graph and on the map. The key point that I want to kind of hit home here is that um you know, when you introduce AI to the public safety community, they're not going to trust it right away. You have to earn that trust. And I

think the best way to be able to earn that trust is to not just respond to their natural language question a relevant response, but to actually show the data visibly, uh, show them that and prove to them that this agent did not hallucinate. You know, here's the data that it is associated with, uh, the response. Um, and it's very tangible and they can interact and interrogate it

as as at will. Um, I think that's how we earn the trust of the the public safety community. And and as as we move forward, um, with more users, uh, hopping on board um, and learn more about the types of questions that they want to be able to ask and and learn about their infrastructure networks and all the different interdependencies there, um, we'll we'll uh, evolve the

the system. Uh, at version one, uh, link, uh, like Kevin had noted, it it is a free-to-use software, uh, soon to be open sourced. And, uh, a key aspect that you you want people to understand about link is that link itself is not storing data. Uh, the data will reside wherever you want it to, and in this case, uh, our DB instance. Um, what link is doing

is, uh, it's it's reaching out to that database and um, identifying based on the the language that you've asked your question, uh, what are the key elements, um, and and developing a a Cypher query to uh, pull out those elements and then answer your question. Um, and so, under the hood, there's a a pretty complex, uh, data model, um, that has, uh, a number of different authoritative

data frameworks, uh, FEMA's community lifelines, the NAICS, um, uh, infrastructure data taxonomy from CISA, and and all of that data has has been organized in such a way that the agent uh understand through the ebbs and flows of different supply chains across all of the different sectors, energy, waste water, waste water, um transportation, telecommunications, so on and so forth. Uh the beating heart uh underneath the the

hood uh for Link is is the agent. Um and so I'm going to hone in and and focus on um on that conversation here. So, next slide, please. Uh from the very start, uh about a year ago, um a little bit less than a year ago, you know, there wasn't a lot across the tech community to to easily build agents, you know. There was some good stuff

out there and a lot of people claimed to have all the answers. but in practice, you know, there there was some challenges and difficulties and uh custom coding these things. Admittedly, you know, the first few iterations of our agents were not so great. Lacking in performance, lacking in quality, be able to uh respond to uh complex uh queries and questions. And and it was hard. Um and

I think uh you know, the day that I discovered that, "Hey, there's a there's a little new feature within Aura that says agents. Hey, I'll I'll I'll check that out." You know, that was a was kind of a game changer for us because it made a lot of things easier. Um you know, as as you develop agents in a custom way, and it's not doing what you

want it to do in the way that you want it to do it, you make slight little changes and and maybe you call that new version of your agent a different name. Um but it's a non-linear process, you know, we we we found as we're iterating that on iteration number four all of a sudden you know you you start to lose quality in in the way that

the answers to questions are are are being exposed to to the user. So with that you know transitioning over to our agents we found that it's a lot easier to iterate because you know it's all templateized and I could really kind of focus in on the signal and filter out all the noise and and focus on just the system systematic instructions and such. Another uh great concept

that I found new and novel I don't know about everybody else but it's our agents separates the actual agent from the agent's tooling. Um and this was kind of a big thing as I discover more about how I want the tools to be developed and have to constantly go back to all the different agents that I had in previous iterations and provide them the new tooling or

or updates to the different tools and that was painstaking and and hard and sometimes you know depending on how that agent was developed. Um you know there'd be little nuances to take care of here and there. Uh but our agents made it simple um we were able to transition a lot of the tools that we used for our semantic similarity search um with our text vector embeddings

um encoded on the purpose and function of all of our infrastructure assets rather than the actual terminology being used and and and and our agents fully understood that and and was able to operate and and interpret that that tooling in the in a good way. Um and you know, we have a a variety of other tools that um perform uh complex queries against the knowledge graph to

be able to pull out good answers for the for our users. And so what glad, you know, uh the team is doing so much good work in this area um because, you know, users ask a whole host of different types of questions and we need uh to be able to have our agents answer those. Uh lastly, the uh slide, please. Lastly, the last point that I wanted

to kind of hit is it's really important for set up structure uh evaluation for your agents. As I mentioned, you know, it's a non-linear process and we want to be able to understand, you know, how each one of the agents that we've built in the past are performing against a number of uh different types of questions. So, we developed a 25-question benchmark um based on that are

representative questions uh based on, you know, the types of questions that we expect from the public safety Um ranging from, you know, operational type questions to analytical type Um and um developing a rubric for how well those uh questions are being answered by different agents. And it's another area that I know um Ed and and team is passionate about. Looking forward to see seeing uh more uh

tools native within our ad to do these evaluations. Um but, you know, today uh it's been we we wouldn't be here as fast as we got um without uh without Ed and his team's help. Um so, that's that's the last point that wanted to make. Uh, if anybody is interested in contacting us, um, uh, on the last slide here, uh, we have our contact information. Oh, uh,

I guess it got moved, but, um, you know, uh, we're looking for, uh, stakeholders and and people interested in in using Link, uh, in production, uh, please, uh, reach >> Thank you very much. Uh, Jesse, and thank you, Kevin, and both of you, uh, for that uh, very thorough and comprehensive, um, use case and how you're using Or Agent. from the world of first responders and emergency

management, we go into the world of improving software engineering organizations. So, Diego, please come on to stage, and let's get started. Diego, you need to click on the button go on stage. There we go. >> Yeah, perfect. Just got it. Okay. So, thank you very much, Ed, for inviting us, and thank you very much for, uh, to to whole team at Neo4j. So, well, my name is

Diego. I'm the founder and CEO at Lean Mot, and, uh, well, we're using the the technology to solve, uh, one of the critical issues that software engineering teams have. So, if you move into the next, uh, So, nowadays, engineering tools don't talk to each other. So, if you think about a how a CTO or a or a head of engineering makes an informed decision today, they have

to go into Jira analytics, into GitLab analytics, cloud analytics, and so on. basically, what we have done is we have um connected all the we have mapped all the data points into um graph database. To solve this issue that um there are 67% delays now in the in the whole in software engineering life cycle. Um we we have understood and we have investigated that there there is

a 3-hour root cause analysis, for example. Um uh this problem costs 2.3 million in a 100 uh software engineering team, 100 uh software engineering developer team. uh next slide, please. So, basically, what we do is we connect all the productivity tools they use, we map the whole workflow, and we detect the bottlenecks that they have in between. So, we uncover from when a ticket is assigned to

someone until the code is in production. And uh we use our agent to analyze all this data with the natural language. Um next slide. So, I don't know if you have um seen the pop culture matrix, but basically, uh organizations are changing, products are changing, and uh what we do is we are more in the process side. So, basically, nowadays, if you want to make an informed

decision or if you uh make an informed decision or if you change any uh part part of your product or service, this will impact directly on your processes. having a clear uh process uh and a clear overview and visibility will help you understand if every decision that you make will impact positively or negatively the the organization. how we use the the Aura agent is um next slide,

please. Yep. So, basically, what we do is we ingest the data. Uh we detect the bottlenecks that they um through natural language. And um for example, one of the questions um So, you can ask single questions as for example, are my senior software engineering teams uh or engineers um overworked uh with junior approvals, PR approvals? And you can get all these data. You can understand if that

problem exists in your team or across team. And uh you can So, we provide the visibility side of it with where you can act upon. So, this is really important because um before we took a a lot of time to understand and to analyze all these And nowadays, uh for example, we collect 1 month of data of a of a client, we found for example, 100,000 relationships

uh between teams, cross teams. And uh we're able to identify what's going on and if the problem exists in in the team or in the organization. uh next slide. Some of the results that we have generated for a client, we took we took a group of 20 engineers and um >> what we have um uncovered is that after 6 months, they were able to increase 135% the

throughput aside from a the use of AI. that resulted in an impact of um So, that they were like delivering like they were as 27 engineers instead of 20. this impacted the organization in um say annual saving revenue of uh $270,000. Um and if you go into next slide, so uh um the the only model that fits our solution and we're looking more in the actional side

of things connecting all these, mapping the whole operational workflow across every tool. And the only uh model that we find uh we found viable the Neo4j graph database connected with the with the Aura agent because it uh provide us with exact exact results and it was amazing um the way we could plan action plans directly on Jira or connected to the different productivity tools. So, yeah, it's

time to um stop using static dashboards and we need to act more into or step into the resolution plans. Thank you very much, Ed. And um yeah, happy to to connect with you all in on LinkedIn or email me directly. >> Thank you very much, Diego. So, let's recap. I'm bringing us home in the last 5 minutes. So, what you've seen today is two very different teams,

uh Napsack Foundation and Lean Mode. They started from a common point. They started from a knowledge graph is stored in their Aura that knowledge graph is the source of truth. It's the data that is in multiple system and all of their relationships. This is what your agent uses to handle uh the questions. So, step two, they create an Aura agent in the console and that helps power

the natural language interactions with your graph. They define tools, test the prompt, validate the responses, all within the Aura console. And finally, [clears throat] uh they integrated their agents inside their UI uh applications using the REST API So, that means that these guys can now focus on crafting user experiences that actually add value to their customers. We hope to have inspired you to learn more about Aura

Agent and try it out in your next project. And if you want to do it, we have a cool way for you to get hands-on. Uh over the next 8 weeks, we'll be running the Aura Agent Hackathon. Taking part is very simple. All you need to do is complete the Graph Academy course, the Aura Agent Graph Academy course, before May 15th. So, you have 4 weeks for

that. The first 500 participants get $100 worth of credits in Aura. Uh next, you build your Aura Agent once you go through the basic and learn the basics. Create an agent. Be creative. Take a data set of interest. Uh uh one that particularly focus focuses lots of relationships. Uh and we'll be looking at, you know, multiple qualities of the agent. So, is is the agent clear? The

the the role? How useful the tool the tools are in in the output ultimately. And once you build your agent, you will have until June 15th to uh tell us about that agent. Uh share it in the uh Neo4j community. And there will be tons of cool prizes on offer. And please check the Neo4j community for all of the details in this Hackathon. But the first step

is complete the course. Learn about Aura Agents. And finally, thank you very much for joining us today. And thanks to our speakers, Kevin Jesse from the Neo4j Foundation and Diego from Leap Motion AI for sharing how they're using Aura Agents in their products today. I guess we can take questions if there are any. don't see any on the session chat. Or manage to get or Q&A Q&A.

Oh, yes, Q&A. To build an agent, do we need to create data or are there already a few data sets? Uh, no, you you do need to create the the Aura agent and the sorry, the Uh, they don't come pre- set with the agent, but there are multiple ways of uh, creating uh, a knowledge graph. If you have unstructured data sources, you can use a Neo4j graph

rack package, for example. Or you can if you have data in relational databases, you can use a data import facility in Aura to, you know, pull that data from um, relational databases and create a create a graph. And you can also use any of the Neo4j drivers to programmatically create a graph. are these Aura credits for people who completed the Academy course? Uh, yes, the credits the

credits will be for the first 500 participants to And I think we're running out of time now, Malcolm, so I'll hand it back to And we'll we'll do our best to uh, answer the the remaining questions in the Q&A offline. So, thank you for joining us. >> [music]