Ontology engineering, done the rigorous way — at AI speed
We combine decades-proven ontology methods with LLM-assisted acceleration. Machines draft; experts decide; tests prove it.
Built on methods that have outlasted every hype cycle
Our process synthesises the best-established ontology engineering methods and adapts them to AI delivery timelines.
The Toronto Virtual Enterprise (TOVE) work introduced competency questions — the questions an ontology must be able to answer — as the way to define its scope and judge its completeness (EIL, University of Toronto). METHONTOLOGY contributed a full development lifecycle, and the scenario-based NeOn methodology put reuse and collaborative, networked ontologies at the centre. Test-driven ontology development closes the loop: questions become executable tests.
TOVE
Motivating scenarios → competency questions → formal terms and axioms.
METHONTOLOGY
Specification, conceptualisation, formalisation, implementation, maintenance.
NeOn
Nine scenarios centred on reusing and re-engineering existing resources.
Test-driven
Every competency question becomes a SPARQL or SHACL test in CI.
Six phases. Clear outputs at each.
Iterative, not waterfall: we cycle through phases per module, so value lands in weeks, not quarters.
Specify
Agree purpose, scope, users, AI use cases and — just as important — what the AI must never see.
Acquire & reuse
Mine schemas, glossaries, documents and SME knowledge; select standards and existing ontologies to import.
Conceptualise
Draft classes, relationships and constraints; review visually with domain experts in short loops.
Formalise & validate
Encode in OWL 2 and SHACL; run reasoners; execute competency-question tests on sample data.
Integrate & govern
Map sources, build the graph, attach sensitivity and purpose policies, wire into retrieval and agents.
Evaluate & evolve
Measure answer quality, leakage and cost against a baseline; version and release through OntologyOps.
Questions first. Classes second.
Vague questions produce vague ontologies. We use a typed catalogue of competency questions so each one has a job: scoping, validating, or testing a design decision.
The typology follows recent research by Keet and Khan, which distinguishes scoping, validating, foundational, relationship and metaproperty questions and maps where each helps in the NeOn process (arXiv 2412.13688). For AI work we add one more family: guardrail questions — "which facts must this agent never receive?"
Don't model what the world already agreed on
Reusing open standards cuts build time, improves interoperability with partners and regulators, and gives you an exit strategy from any vendor — including us.
Financial services
FIBO, LEI, ISO 20022 concepts
Insurance
ACORD data model concepts, FIBO parties
Healthcare
HL7 FHIR, SNOMED CT, LOINC, RxNorm
Manufacturing
ISA-95, Asset Administration Shell
Security
STIX 2.1, MITRE ATT&CK, D3FEND, UCO
Privacy & policy
W3C DPV, ODRL, PROV-O
Cross-domain
schema.org, SKOS, OWL-Time, GeoSPARQL
Upper level
BFO or gUFO where rigour pays off
AI accelerates the build. Experts own the meaning.
What the models do
- Extract candidate terms and relations from schemas, PDFs and tickets
- Propose definitions, synonyms and alignments to standards
- Draft competency questions from user stories
- Generate test data and SPARQL for CQ tests
- Flag likely sensitive properties for review
What the experts decide
- Which concepts exist and what they mean
- Modelling patterns and upper-level commitments
- Sensitivity classification and purpose rules
- Acceptance of every alignment and axiom
- Release, deprecation and change policy
How we make design decisions
Questions before classes
If no competency question needs it, it doesn't go in.
Small core, modular edges
A narrow, high-confidence core extended iteratively beats a grand unified model.
Policies are data
Access rules live in the ontology, versioned and tested like everything else.
Evidence or it didn't happen
Every fact keeps its provenance; every answer shows its path.
Open standards, no lock-in
OWL, SHACL, SPARQL and PROV run on any compliant store.
Your team owns it
Pairing from day one; documentation and training are deliverables, not extras.