Tables with an RDF dress
There are two ways enterprises build an ontology and Stéphane Fellah thinks both of them are wrong. Top-down from a philosophical upper ontology buries the project in arguments about what a quality is. Bottom-up from your existing schemas re-imports every application-centric assumption you were trying to escape, and he has seen where that lands: 15,000 classes and properties that are, in his words, just tables with an RDF dress. His alternative starts somewhere less obvious, with the decision questions the business actually needs answered, which scope the model and double as unit tests. He closes on why agentic AI will not work until agents can describe themselves semantically and present credentials before they act.
This episode hasn't aired yet — get notified the moment it drops.
Stéphane Fellah
Geospatial semantics pioneer who entered the field through OGC standards in the late 1990s and now builds knowledge-graph-grounded AI, including his census.market work on verifiable agent assertions.
What we'll cover
- 01The two ways enterprises build ontologies that both fail: philosophical top-down (BFO, CCO) and bottom-up from existing schemas, and why the bottom-up version imports every assumption you were trying to escape.
- 02Competency questions as the unit of scope: how a decision question bounds a module, and how a SPARQL query turns it into an actual unit test.
- 03The three-layer build, cross-domain foundations (W3C ORG, SSN, DCAT), a business layer, then the vertical, and why small ontologies with few classes are the ones that survive.
- 04What an LLM-generated ontology looks like when nobody checks it: 15,000 classes and properties, a relational model in RDF clothing, zero interoperability.
- 05Why 80% of an ontologist's work is now automated and the remaining 20% is the whole job, plus why he stopped using Protégé and TopBraid.
- 06Implicit semantics as the actual root cause of interoperability failure: a lat/long column with no projection or unit, and the Mars orbiter that crashed over kilometres versus miles.
- 07Don't migrate your data, lift it. Adapters that turn API JSON into a graph, and why MCP returning JSON still leaves you with the same problem.
- 08A concrete learning path for web developers: JSON-LD, then RDF, RDFS, OWL, SHACL, SKOS, and where to look at open linked data in the wild.
- 09The closing argument: agentic AI needs a trust layer where agents present verifiable credentials and satisfy a contract before they act.
Related episodes
An ontology fits on a Post-it note
Casey Hart is a rare thing: an actual ontologist. A philosophy PhD who answered a job ad from Cycorp and spent a decade building knowledge for machines under Doug Lenat and then at Olive, Amazon, Gro Intelligence, and Ford, he spends this episode deflating the word everyone is suddenly selling. An ontology, he argues, is just a summary of what your business cares about and how those things relate — you can start one on a Post-it note. He separates the machine-learning "system one" from the deterministic "system two" that ontologies supply, makes the case for a hybrid, and walks through building one from the ground up: taxonomies, relationships, turtle files and triple stores — or just the metadata, so you get value before migrating a single row. Along the way: why "hallucination" flatters a text generator doing exactly what it was built to do, the open-world versus closed-world assumption, and why vibe-coding an ontology out of an LLM is a fine way in but not a finished asset.
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.
Every silo is its own context
Dave McComb has been building enterprise systems for fifty years and running Semantic Arts for twenty-five. His rule of thumb is one application for every ten employees, so a hundred-thousand-person company is running ten thousand databases that each name, structure and identify everything differently. That is also his answer to the age of context: the problem was never a shortage of it, it is that every silo has its own. Along the way he explains why an LLM will hand you an ontology that opens cleanly in Protégé and looks like exactly what you asked for, why that is the trap rather than the win, and what he found when he put semantic lenses back on his own accounting system after twenty years of not looking at it.