December 11, 2025

My Top 5 Favourite Talks from NODES 2025

Christian Miles Graph Community AI, cybersecurity, Cypher, graph databases, graph technology, healthcare, knowledge graph, Neo4j, NODES

At 36 hours long with over 130 different talks across too many time zones to count, the Neo4j Online Developer Expo & Summit – NODES 2025 – was as enriching as it was interesting. This year’s event featured conference tracks organized by region, making it one of the easiest conferences to attend online from anywhere in the world.

I (Christian Miles) didn’t get a chance to attend all of the talks live, but the Neo4j team has done the graph technology world a massive favour and published all the NODES 2025 talks and videos on their Youtube Channel.

With the darker evenings earlier on this side of the world I spent a bit of time sitting down with a cup of tea to explore some of the talks I wasn’t able to attend. Frankly, I think there’s some real gold in these talks, so I thought I’d do my part to shine a light on ones you might have missed. Kudos to the Neo4j team for finding and organizing such amazing speakers for this year’s event.

Without further ado – and in no particular order – here are my five favourite talks from NODES 2025.

The Company You Keep: Mapping Scientific Collaboration with Graphs

by Pedro Sader Azevedo

Pedro Sader Azevedo presents this accessible talk on extracting hidden collaboration structures from networks of scientific papers, authors, and institutions.

Azevedo was looking to test a hypothesis for his final thesis at Unicamp: Does a researcher’s first external collaboration improve their citation rate? The idea here is that once a researcher collaborates with people outside their home institution (and outside their usual cluster of collaborators), they gain exposure to new networks of researchers who might cite their work.

The implementation approach in Neo4j is to take the original graph structure and then project downwards to an author-only graph by adding relationships between co-authors. After running a community detection graph algorithm to find clusters of heavy collaboration, use a definition of “external collaboration” to understand how citations change over time.

I like this talk for three reasons:

Citation networks are really dense and interesting. Whenever I read a paper I end up with 10 tabs of PDFs to read later.
The result is weird! The control group actually outperformed the experimental group!
It’s a nice heartwarming ending as Azevedo reflects on completing his degree and attending the graduation ceremony with his family. It’s a nice reminder that it’s not all about the stuff that happens on the computer screen.

Graph-Powered Job Matcher: From SQL Tables to Neo4j RAG in Production

by Otávio Calaça Xavier

To me this is the perfect talk. It’s accessible to the audience without hiding the complexity and interesting parts of the implementation.

Otávio Calaça Xavier walks through the design of a system to help recruiters and hiring managers find the perfect candidates for their open roles.

Data used in recruitment technology has the same pitfalls that we see elsewhere. Siloed data sources, complex retrieval requirements, and nuanced real-world considerations that could be problematic if over-simplified. Xavier walks through the options for loading the data sources into Neo4j, describes his AI-enhanced architecture with very little hype and gives concrete examples and benefits of his approach.

Xavier points out the ability to use Neo4j to construct a path through the graph database to avoid opaque results and actually explain the outcome from the search.

I enjoyed the example of parsing job descriptions to determine must-have skills, nice skills, country – and the fact Xavier calls out that you can use traditional NLP tools to do this job reliably, efficiently (and cheaply!). I also appreciated Xavier’s breakdown of costs for the implementation as this is an important factor that presenters often avoid.

My main takeaway: Smart graph data modelling goes a long way when solving messy real-world problems.

Early Path Traversal Pruning

by Jan Žák

In this talk, veteran graph engineer Jan Žák walks through the new allReduce function in Neo4j that allows users to express complex path conditions evaluated and pruned during the traversal, rather than post-filtering.

It’s a short and sweet presentation: Zak presents the core options to prune complex path queries, introduces the allReduce function syntax and shows off two concrete examples of it in action.

That’s it! That’s the whole talk!

From Side Project to Startup: Graph Lessons from the Cartography Journey

by Alex Chantavy

Everyone loves a tale of the hero’s journey, one where you root for the protagonist as you see them overcome challenges, and cheer from the sidelines. In this talk from Alex Chantavy, we get introduced to Cartography which was created as an internal tool at Lyft and then Chantavy’s decision to start a new company to bring the lessons learned to the industry as a YC-backed company.

Working on Microsoft’s red team (simulating cyber adversaries), Alex found that graph databases were essential for modelling cloud cybersecurity relationships – in particular tracing attack paths from assets on the internet along authorisation paths to sensitive systems containing customer data. Later, at Lyft he worked on the team that open-sourced Cartography, a Python tool to ingest data from the main cloud providers into Neo4j for analysis.

Building Cartography wasn’t without challenges. It was interesting to learn more about how the team worked hard to scale Neo4j to meet their performance requirements. Chantavy also explains how an ORM mapping approach helped abstract away some of the complexities of Cypher to bring business logic querying without needing to learn another query language.

As part of his talk, Chantavy cites the evergreen quote from John Lambert: “Defenders think in lists, attackers think in graphs.”

For my part, I talk with lots of companies using G.V() to visualize attack graph paths through various complex network topologies and infrastructure. Chantavy notes that a real breakthrough was the use of the tool for vulnerability management for complicated container image lineage.

Graphing Clinical Data for Smarter AI Systems

FHIR in the W.H.O.L.E.: Graphing Clinical Data for Smarter AI Systems

by Krishnendu Dasgupta

Good talks give you a glimpse into a world that you otherwise have no exposure or understanding. The best talks leave you with a new understanding of acronyms that you can use in the future to make you sound smarter.

‘FHIR in the W.H.O.L.E.: Graphing Clinical Data for Smarter AI Systems’ presented by Krishnendu Dasgupta shows an interesting approach to providing actionable semantic context to data shared using the healthcare standard FHIR (Fast Healthcare Interoperability Resources 🤓, pronounced ‘fire.’)

The WHOLE framework transforms flat FHIR records into rich knowledge graphs where patients, conditions, medications, encounters, and temporal metadata are interconnected nodes. Complex queries can then be executed against the knowledge graphs to understand the patient journey.

There’s a human-centric angle to this use of AI: By reformulating the data source away from being ‘resource-focused’ and more ‘patient-focused’, the framework captures the complexity of individual patient journeys and helps build notable clusters of similar experiences.

The use of 7-8B parameter open weight models like Qwen adds another advantage: a privacy-first architecture well-suited to handling the sensitive data central to the system.

Dasgupta also suggests that this model can be used to improve and streamline automated matches for patients to be recommended for clinical trials by parsing natural language requirements and identifying those who are a perfect fit. I’m interested to learn more about the performance of this approach over existing techniques, it seems like the perfect blend of knowledge graphs and modern AI stack that could ultimately improve and maybe even save lives.

Runners Up & Honourable Mentions

Honestly, picking just five talks from NODES 2025 was very difficult. Here’s a quickfire list of a few more that I enjoyed:

Conquering Neo4j Errors: GQLSTATUS Codes to the Rescue. A conference talk about error codes?! Not all talks can be about fancy new AI techniques. Anyone who’s worked with databases knows that error codes are important.
Taming the Evolution of Cypher: What Cypher Versions Mean for You. This talk is a good summary of the split between Cypher 5 and Cypher 25. At G.V(), we care a lot about query languages, and this is a great deep dive.
Building an Entire Graph Application with Nodestream. A great talk about the declarative Nodestream framework and how to maintain your graph application on day two and beyond. For more on using Nodestream to create your graph, check out this quick tutorial video by my colleague Amber Lennox.

Looking Ahead to 2026

As we approach the end of the year, it’s time to start looking forward to graph tech conferences and events in 2026.

In January, the G.V() team will be presenting at Data Day Texas, and in May, we’ll be at the Knowledge Graph Conference in New York City. If you’ll be attending either one, reach out and let me know.

I hope you’ve enjoyed these NODES talks above as much as I have. The NODES conference is a winning formula so it’s not a surprise to see the Neo4j team already planning for the next one with NODES AI coming up in April – see you there!

Building a graph-powered application with Neo4j? Try out G.V() and amp up your Cypher queries, data model tracking, graph visualization, and data exploration – all in a single IDE built for developers.

Let's Go