Graphs and AI: Unlocking Real Business Value

The race to implement AI is leaving many businesses struggling with hallucinations, compliance concerns, and disappointing results. What's missing? According to Dominik Tomičević founder and CEO of Memgraph, it's the critical connection between AI and graph technology.

Dominik shares his fascinating journey from winning a Bill Gates grant at Davos to founding Memgraph after encountering scaling walls with traditional graph databases. His team has pioneered an approach called GraphRAG (Graph Retrieval Augmented Generation), which grounds large language models in living knowledge graphs rather than static document collections. The result? Fewer hallucinations, traceable reasoning, and faster access to institutional knowledge.

Real-world applications are already showing impressive results. NASA's People Knowledge Graph has cut discovery time in half by helping teams find expertise and relationships across the agency. Health researchers at Cedars-Sinai are using knowledge-driven AI to support Alzheimer's research. Financial institutions are identifying fraud rings by analyzing network behavior rather than relying on easily-circumvented rules.

What makes graph databases so powerful for AI applications? Unlike traditional approaches that prompt models with disconnected documents, GraphRAG provides linked, summarized facts and multi-hop paths that justify answers. This matters tremendously for enterprise adoption, where trust, compliance, and regulatory requirements demand explainable AI with clear audit trails.

Looking ahead, Memgraph is focused on making GraphRAG more accessible with turnkey solutions that combine vector and graph capabilities in one engine, better observability tools, and deeper integrations with popular frameworks. For businesses navigating the AI landscape, the message is clear: LLMs are just the finishing layer – you need the right data architecture underneath to deliver real business value.

Ready to explore how graph-powered intelligence could transform your organization? Memgraph 3.0 provides the building blocks to get started on your journey to more reliable, explainable AI.

Everyday AI: Your daily guide to grown with Generative AI
Can't keep up with AI? We've got you. Everyday AI helps you keep up and get ahead.

Listen on: Apple Podcasts   Spotify

Transcript

Speaker 1 0:02

Hey everybody , fascinating discussion today as we talk about unlocking real business value with graph-powered intelligence and what it takes not to get left behind with AI these days . Dominic , how are you ? I'm pretty good , how are you , but maybe next year ? Before that , maybe start with your story about founding Memgraph and what was the big idea , the vision behind the company .

Speaker 2 0:31

Awesome , let's dive right in . So I'm Dominic . I'm one of the founders , I'm currently CEO of Memgraph . I'm a techie computer scientist founded Memgraph about 10 years ago . We've been in stealth for a long time , kind of after that , building the whole product and then building the company in public for the last couple of years .

Speaker 2 0:54

It's an interesting story how it came into graphs . I think it came into graphs the hard way by hitting scaling walls . A decade ago I was building a system that needed to reason over fast-changing , highly connected data , and the available graph technology was great for static analytics , but it really struggled with high-throughput updates and millisecond queries . My co-founder , marco , and I kind of teamed up on that issue and that led us to build Memgraph as an in-memory graph database that's built for real-time workloads and more recently , graphrag , which has been a big use case for us , where kind of LLMs are grounded in living knowledge graph rather than static snapshot . Knowledge graph rather than static snapshot . Today that shows up in customer work , like NASA's People Knowledge Graph , where teams discover expertise and relationships across the agency , and health research , for example , at Cedars-Sinai , where knowledge-driven AI supports Alzheimer's research . So it's been a good ride so far .

Speaker 1 2:01

Well , so many important use cases and we're hearing AI and LLMs in the news every day , but why is the understanding and leveraging graph tech also really critical for staying competitive ?

Speaker 2 2:21

So I think the key term is graph RAG , which stands for Graph Retrieval , augmented Generation , so it blends traditional RAG with the knowledge graph . So instead of asking LLM to remember everything , you essentially extract entities and relations from your corpus , you build a graph and then at query time you retrieve both semantically similar content , which is through vectors , and also structurally relevant subs which you kind of use the graph search for to find multi-hop paths , constraints , stuff like that . The original Microsoft research paper on GraphRag that work showed that graph-enriched retrieval boosts kind of quality on complex narrative datasets because the model is prompted with linked , summarized facts rather than a bag of documents and the outcome is kind of fierce hallucinations in answers and also you can get a traceable rationale . So the upside is real richer reasoning , fresher answers . So no retraining to reflect new facts , explainability via graph paths .

Speaker 2 3:20

So the counterpoint is of course that GraphRack isn't plug and play , so it's not simple as the vector search . So you must invest in data modeling , entity resolution and graph maintenance , and our approach at Memgraph has always been to lower that particular barrier . So have the vector search and GraphRack in one system , templates for all the GraphRack pipelines , and teams can ship business value without kind of months of plumbing .

Speaker 1 3:49

Amazing . So you know you've been talking about combining graph data with AI for a long time now . You mentioned 10-year journey . What was the aha moment when you know companies finally get the power of that combination ?

Speaker 2 4:04

Yeah , it's been always an uphill battle kind of , I think , from the start , because you either don't get the graph or you're a real , true fan and everything in the world seems like a graph . And I think it's been a couple of years ago when the whole LM thing started , when we're like , oh , I mean , the grounding knowledge must be really important , so there must be a big use for graph . And I think with some of the research papers that have been coming out and some of the terminology being more discussed and literally everyone kind of watching the topic of AI , so I think that was kind of the pivotal moment where more people started paying attention . The question is how many production GraphRack systems are there real production ? We know some of the good ones that are built with Memgraph , but a lot of people in our community are exploring and I would say that 80% of the discussions that we see in the community these days are always kind of about AI at least , and people are just trying to kind of catch up and get on board with the graph .

Speaker 1 5:15

Fantastic , and Chachi Pete told me you were personally awarded a grant by Bill Gates back in 2011 . That's pretty impressive , not something a founder can usually drop into a conversation . Was that a hallucination or was that actually the case ?

Speaker 2 5:34

It's a good point . Yeah , I actually have . It's been an amazing experience . So the Imagine Cup Grants program it was announced in 2011 . So we were the first year winners and the inaugural grants were awarded at the World Economic Forum in Davos . That was January 27th , I think , and 2012 , which is a long time ago . So I was a captain of my kind of creation software development team that competed on Microsoft's Imagine Cup competition and the top finalists worldwide finalists , were invited to participate in the Imagine Cup grants for program . So we were one of the four winners , alongside the teams from the US , Jordan and Ecuador . And yeah , so we had a project that was very data-driven was called Kidnect .

Speaker 2 6:31

If anyone remembers the Kinect-based motion tracking in gameplay for Xbox . We essentially used the motion tracking to support children's physical therapy programs , so we kind of hacked the Kinect , connected it to the computer . Microsoft was kindly kind enough to allow us to do that , so that was that was awesome . So event was a part of the broader Microsoft initiative . I think it was championed by Bill Gates to help student projects become real ventures . So , yeah , so there's a there's . There's a fun bit of lore here . I think People often say a word about Bill Gates , which reflects the whole kind of his visible role around the program Of course , davos activities , around the winners and everything else . But I think the authoritative kind of record is the participation and kind of the project development , what we've been doing with our project Kinect , and also kind of highlighting and meeting founders from other four different teams . So yeah , I think that experience shaped how I think about turning ambitious prototypes into production systems , which is a straight line from there to building Memgraph .

Speaker 1 7:45

Wonderful , what a great story , and you mentioned a few anecdotes stories on leveraging Memgraph . Can you walk us through a kind of real world example on how a business can or has leveraged Memgraph and GraphRag to get real business results ?

Speaker 2 8:12

immediate examples that comes to mind is workforce or expert finding assistance . So NASA built a people knowledge graph that combines people , projects , skills , essentially decades worth of project documentation and research and everything else you can kind of muster up together and build a comprehensive knowledge base . So they layered LLMs on top . To answer questions like who's built similar projects , who's worked on similar missions that we have worked on on X subsystems , stuff like that , the LLM always generates the final narrative answer . But the thing is the graph supplies the verified relationships and multi-hop context . That is kind of . They dug it up across data silos across the entire organization and structured it in a graph . So in practice that halves the discovery time . It scales institutional knowledge beyond those silos where data has resided . We see similar wins in , I think , customer support , risk analysis .

Speaker 2 9:15

So we have other public stories with Microchip Sayari about cutting latency , surfacing deeper connections , and most of our customers are building systems that have some connection with AI , even if previously they didn't get kind of with Memgraph because of AI . I think the point is how to implement it quickly . So you ingest your documents , you run the name identity recognition as relationship extraction , you build a graph , you index with vectors and graph and then use kind of our tooling to wire graph rack chain on top . It expands the subgraph around the question , summarizes with citations and returns paths that justify the answer . So our 3.0 release kind of ships with those building blocks and we have a lot of other things in the pipeline that we're kind of building to make this whole thing even easier .

(Cont.) Graphs and AI: Unlocking Real Business Value

Speaker 2 10:13

Fantastic .

Speaker 1 10:16

You've also warned us to be you know careful about leveraging LLMs in the enterprise . What's behind your words of caution ? And you know how can we get more safety ? Let get more safety for real business applications .

Speaker 2 10:32

Yeah , I think I'm kind of very bullish on the process and the whole progress and everything that's been happening . But the problem is always kind of hallucinations and leaking kind of sensitive information . Trust and compliance is a big thing , it's non-negotiable so , and even the EU AI Act has started kind of phasing in obligations transparency for general purpose AI , more stringent rules for high risk systems . I think there's some key milestones set as well . So if your AI can't explain itself with kind of source grounding , reasoning steps , data lineage , so your face operational risk , regulatory friction , so graph-grounded architecture can help you there , can help you make it much easier to produce the audit trails and evidence that regulators expect .

Speaker 2 11:28

But it's also kind of time to insight cost . The constantly fine-tuning retraining models to keep up with fast-changing Business data is expensive and slow . Rag , especially GraphRAG , lets you update the database and not the model , so answers reflect the latest reality without the retraining cycle , which can be expensive . So a lot of companies that are implementing and building kind of these systems for their customers , their guidance is that for many enterprise use cases , retrieval is the faster , cheaper lever that you can have , while fine-tuning can be reserved for style , format or genuinely new capabilities . So there's lots of kind of gotchas kind of with current LLM systems .

Speaker 1 12:16

Yeah , lots of roadblocks to consider there . So where do you see AI-powered graph databases heading over the next couple of years ? What's the next big unlock that you envision ?

Speaker 2 12:29

Yeah . So we've been thinking kind of long and hard on this problem . We talk about it every day . The first two , I think , priorities that we have . The first one is to make GraphRack turnkey , so vector plus graph in one engine , agentic retrieval workflow , so the system kind of chooses the right retrieval and ranking strategy per query . So you don't have to do a lot of the heavy lifting , you can kind of focus on the fine tuning and kind of learn as you go . But you have a complete system as soon as possible and also better observability on why the answer was produced . Because once you get scorched by some hallucinations , and especially kind of if you really were confident that the answer is right , it's like , oh shit , it's clearly wrong . So then you start kind of being very cautious about whatever kind of comes next . So you need good observability and you need to know kind of how the answer was produced so you can gain confidence in the answer . So our 3.0 release that we did earlier in the year , it shipped kind of the big primitive . So you can kind of build those things yourself right now . But now we're expanding the tooling and templates so customers can kind of go from data to production chatbots and agents with grounded answers much faster .

Speaker 2 13:52

I think the second thing for us keep investing in developer velocity , enterprise hardening , which is always critical for making sure that what's kind of private keeps private . So our 3.0 Memgraph Lab release also updated a lot of those things and it's shortening the path from schema design to explainable AI in production . We have file-based kind of locking and controls . You can really pinpoint what LLM has access to based on who the person is asking the question . So you essentially stay compliant as an enterprise . But I'm also very excited about the kind of the developer progress in terms of the developer velocity , where you'll see kind of deeper integrations , which is frameworks , evaluation tools , more guidance and just documentation , more resources you can use to get kind of from reference architectures to live deployments in various sectors . So we have space , healthcare , finance , energy . So leveraging a lot of the knowledge that our customers have acquired and making it available for developers around the world .

Speaker 1 15:15

Fantastic and that's sorely needed . We saw some recent data that a lot of enterprises are kind of slow to get their production systems to market with real , tangible ROI . What are some of the mistakes maybe misconceptions they have about AI or graph tech from the enterprise point of view ?

Speaker 2 15:35

Oh , I think there's quite a few . The thing that comes to mind is that just putting everything in an LLM is going to magically make things work .

Speaker 2 15:46

But LLMs are a really good finishing layer , but the whole stack needs to be much better . So data needs to be accessible , it needs to be cleaned , it needs to be authoritative , and then you need really good retrieval systems , graphrack being one of them . You need a good kind of vector tool as well and you need to kind of connect the whole stack . And if you want to build anything that goes further from just questions and answers , you need to think about the memory , so agent memory . If you want to build agents and workflows , you need to think about episodic reasoning how the task you're trying to kind of accomplish , what are the instructions ? How do you train your agent to understand what the task should be and how it should be executed ?

Speaker 2 16:40

So a lot of the things need to be right in order to get close to what humans are capable of doing . But I think the messages start from wherever you have the lowest hanging fruit , wherever you have something that can be automated quickly . So you gain kind of experience and you kind of see those shortfalls by yourself kind of in the organization and you can really think where to focus more on the architectural side , so kind of , instead of just pushing the top down or the bottom up approach , kind of combine approaches and and really figure out where it makes sense to invest time yeah , challenging , but so much room for innovation .

Speaker 1 17:25

Any unexpected use cases that you were surprised by with your customers or thought , wow , I didn't imagine that could be done with mammograms .

Speaker 2 17:37

Yeah , I think there's always someone doing something crazy with graphs . There's lots of use cases , I think , in cybersecurity . There's traditional ones from kind of identity and access management , but real-time threat detection in automated systems . This is going to be greatly improved by AI and GraphRag workflow . So you essentially have millions of new data points that are being created and it's hard for any human to keep track of what's going on . So you need automated systems .

Speaker 2 18:10

But a lot of the I think automated systems previously have really focused on specific problems that have been kind of identified and found . But then when you apply it more broadly to agents and you're trying to reason about connected events , it really kind of brings it all together . You can surface the right alerts , you can prioritize them , you can make sure that from all of the kind of forest you essentially see the relevant things that you need . So it's going to happen the same in the observability space in the network . If you're building like your cloud resources or your big company have a lot of networking equipment and devices , have a lot of networking equipment and devices and even if you're kind of going beyond , so on a single instance it's not even . I have a server , so now my server has multiple virtual machines . It has every virtual machine can have containers . I have Kubernetes , clusters , so the complexity is just rising exponentially .

Speaker 2 19:19

So I think a lot of the potential kind of in the digital cybersecurity space and that's , I think , where I've seen the largest data sets , which is always more challenging to kind of surface insight . So naturally I kind of tend to gravitate towards much , much harder problems . But similar things exist in finance , for example . So tracking fraud rings and really making sure that fraudsters , who are getting smarter every day they're circumventing the traditional rules . They really need to observe the network behavior , how they behave together .

Speaker 2 19:58

Who are the people who are colluding together to execute those crimes , and even if you have your own bad actors inside the organization who are aiding them to be able to avoid the kind of rules , and that's where you kind of need that top level and that's where you kind of need that top level view of the graph . So yeah , I think just a lot of things are happening in so many different industries . It's really tough to generalize all of the solutions that people are creating . But I think that kind of speaks to the flexibility of the graph structure and what you can actually put in it and how many different problems you can solve . Which is so you can implement graphs for one particular use case that you have in your organization , that you can grow to solve a multitude of problems .

Speaker 1 20:57

Incredible . Well , what an amazing opportunity you have ahead of you . Just a final thought Many of us know of your home country , Croatia , as a great vacation destination , but it's also becoming a European global tech hub . You know , talk about the tech landscape a little bit and what an amazing place it is . To start a company or invest or acquire companies really is the next big thing .

Speaker 2 21:24

Yeah , it definitely evolved a lot since we started Back . When we started there , there was only a few business angels . We didn't have VC funds . Now we have a lot of them who are operating in the area local funds , but the global funding landscape has also expanded heavily , so there's a lot of capital available . There's a lot of interesting companies being started .

Speaker 2 21:51

We've seen the transition from going from kind of a consulting-based industry where when I was in college 15 years ago , most of the tech companies sort of tech companies were software development houses and there's nothing wrong with that . But it's just interesting how we're also transitioning to highly scalable SaaS products and product-based companies . Saas products and product-based companies I mean you know all of the winners . You know Infobip , you know Rimats Electricars . We have so many world records they have won with that . So there's lots of expertise and a lot of people who have forged their careers kind of in the early days of those companies are now starting their own companies . So things are expanding quite a bit . Zagreb is a big hub . There's a lot of people , there's a lot of talent , we have some great universities there as well , and the coast is a big benefit . It's two or three hours by car and it's very easy , very easy to get to and to relax after a kind of hard week worth of work . So I would encourage everyone to visit All right .

Speaker 1 23:09

Well , that's something I would totally agree with . Thanks so much for joining and sharing the mission and the vision , onwards and upwards .

Speaker 2 23:17

Thank you for having me . I'm excited about the future .

Speaker 1 23:20

Thank you and thanks everyone for listening , watching and sharing this episode , and be sure to check out the TV show at techimpacttv on Bloomberg and Fox Business . Thanks , dominic , thanks everyone , thank you .