Bring the graph to the data
A citizen in Italy asks a public chatbot when their civil-service exam is. The honest answer keeps moving: dates get corrected, sessions get cancelled, and amendments pile up across official notices. Ask a language model alone and it answers confidently, sometimes with a date it picked at random, sometimes with a session that no longer exists. Giuseppe Futia built the knowledge graph that keeps the answer current. It does the deterministic cross-document work an LLM cannot be trusted with, matching each amendment to the exam it modifies, superseding old versions, cascading cancellations, and staying auditable throughout, inside a production pipeline behind a chatbot he says serves close to a million citizens. A former La Stampa journalist with a PhD from Politecnico di Torino, and the first European guest on the show, he also opens up the regulated side of his work: healthcare data that cannot leave the country, inference that runs on premise, and patient records he is not allowed to move even inside his own infrastructure. His portable lesson is the entry point. Prove value on one well-defined task the institution already needs, then earn the right to expand.
This episode hasn't aired yet — get notified the moment it drops.
Giuseppe Futia
Builds graph-based AI systems at CSI Piemonte, the public IT company serving Italy's Piedmont regional government. He holds a PhD in Computer and Control Engineering from Politecnico di Torino, co-authored Knowledge Graphs and LLMs in Action for Manning, and runs a nine-week cohort on knowledge-graph engineering at kg-training.com. Before any of the engineering he spent seven years as a journalist at La Stampa, one of Italy's oldest newspapers, and he worked on knowledge-graph projects at GraphAware, including the Rockefeller Archive Center graph.
What we'll cover
- 01Seven years at La Stampa before the PhD: what reconciling conflicting sources in a newsroom taught him about structuring knowledge for machines.
- 02The system that is actually running: a public chatbot for Italy's public-sector job competitions, and the graph pipeline that builds its knowledge base.
- 03One amendment, end to end: how a new notice gets matched to the exam it modifies, how the old version is superseded, and how a cancellation cascades to dependent sessions.
- 04The failure modes that ruled out the model on its own: arbitrary date selection, cancelled sessions resurfacing, and reconciliations it invented.
- 05Why established institutions are the hard case for a new data paradigm, and the entry-point strategy that got his graph into production: prove value on one well-defined task, then expand.
- 06The regulated work next door: healthcare data that cannot leave the organization or the country, and why the big model APIs are off the table on day one.
- 07Batch over latency: on-premise inference with vLLM for annotating clinical narratives against an existing ontology, at volume, under a deadline.
- 08Patient records he cannot move even inside his own walls, and the Neo4j virtual access that materializes data into the graph only at query time.
- 09The honest line between the live competitions pipeline and the healthcare GraphRAG proof of concept, and what separates a demo from a deployment.
- 10The CFO escalation: why not just put every notice into a long-context model, and whether there is a churn rate below which the graph is genuinely overkill.
Related episodes
Relationships should be the product
Himanshu Singh has built knowledge graphs three times at three very different scales: a politics subgraph inside Microsoft's Satori, a zero-to-one product graph at eBay, and now Netflix's Entertainment Knowledge Graph, where he leads engineering. The line he keeps returning to is that relationships should be the product. Node count is not the measure, and a graph that duplicates what already lives in your CRM or your warehouse is mostly cost. He is notably relaxed about technology choice, pointing out that Netflix built its own real-time graph abstraction over a key-value store because no native graph database could absorb their write volume. What he is not relaxed about is data quality at the point of entry, because once bad data is in a graph and connected to everything else, undoing it is very hard.
Everything was associated with everything else
Mike Dillinger built the Economic Graph at LinkedIn, worked on knowledge graphs at eBay, and now runs hypergraf.ai, which he describes as an X-ray machine for other people's graphs. He tells the story of an organisation that spent millions building out "associated with" relations until everything in the graph was, one way or another, associated with everything else, and the query expansion got so robust it undermined the system it was built to serve. Underneath the story is a linguist's diagnosis of why these systems fail: language models operate on strings, and a string is not a concept. Query canine, Barker, Pooch and Puppy and you get four different answers. His conclusion is that the human check everyone has pushed downstream of the output belongs upstream, on the knowledge foundation.
A different kind of answer, not a better number
Karthik Soman invented KG-RAG — knowledge-graph-based retrieval augmented generation — while building biomedical knowledge graphs at UCSF, and now leads enterprise-scale agentic AI at SAP America. One question carries from a PhD in computational neuroscience to the enterprise: how do you build intelligent systems that actually work in the real world? His answer isn't a better accuracy number but a different kind of answer — one a human can trace, question, and learn from. He walks through the case that convinced him: enriching patient records with a 40-million-node biomedical graph surfaced an olfactory-receptor gene that flagged Parkinson's five years early, catching a prodromal case a clinician had missed — not because the model was more accurate, but because it pointed at a mechanism. Then he moves to the enterprise, where the curated ontologies of biomedicine don't exist. You lean on the topology already inside your documents. Graph-based reasoning turns out to be a sixty-year-old idea that LLMs merely made usable on the fly. And a graph earns its keep over vector RAG in specific places — multi-hop questions, smaller models, private data the model never saw — before the least glamorous advice in AI: data hygiene first, then AI hygiene.