Skip to main content

Service: ontology engineering and knowledge graphs

Ontology engineering in RDF, OWL 2 and SHACL

We design domain ontologies, build knowledge graphs, and prove they are correct rather than merely well formed. Semantic web work in the full stack: RDF and linked data, OWL 2 and SKOS modelling, SHACL constraints, SPARQL and triple stores, taxonomy design and GraphRAG retrieval. We built the ontology extension tool for the UK National Digital Twin Programme, and we have published 51 open ontology studies across 29 public repositories, every headline computed at least two independent ways. If a hypothesis died during the work, the study says so.

What we deliver

Most ontology projects fail in the same place. The model is built, it loads, a validator reports zero violations, and the programme concludes the data is trustworthy. It is not. SHACL, the W3C validation language everyone reaches for, operates under the open-world assumption: it silently passes any term it has no shape for. We measured this. Across three real vocabularies, open-world SHACL validated as conformant every one of 300 data graphs carrying a term that does not exist, drawn from 418 fabricated terms. A closed-world vocabulary gate caught all 300 with no false positives on clean data. That is not an edge case. It is the exact failure mode of a language model asked to author RDF, which is how a growing share of knowledge graphs now get built.

So our engagements ship a check that can fail. We work in OWL 2, SKOS and SHACL, on top of standard triple stores, and we add a closed-world gate over the vocabulary so a hallucinated or drifted term is rejected instead of accepted in silence. The tooling is our own open-source engine, Open Ontologies, a single Rust binary with an in-memory Oxigraph store, a native OWL 2 DL tableaux reasoner, SHACL validation and SPARQL. There is no JVM to operate and nothing leaves your network.

A typical engagement runs in four steps. We read the source standards by hand rather than trusting their documentation, because that is where the defects are: we have reported faults in the IATA ONE Record data model, in FIBO, in the W3C JSON-LD specification, in DCAT-US at the General Services Administration and in the USDA National Agricultural Library thesaurus, and each one is a public thread you can read. We then model the domain in OWL 2 with SKOS for the controlled vocabulary and SHACL for the constraints that have to hold. We build the instance graph from the real source data, not a sample. Finally we compute every headline figure two independent ways, using different libraries, and reconcile the difference before anything ships.

We also train teams to do this themselves. Our ontology engineering course covers RDF, RDFS, OWL 2 TBox and ABox modelling, SPARQL, SHACL, triple stores, GraphRAG and ontology governance across 15 lessons, and it is free. Public sector training is procurable through CCS RM6219.

How this differs from the usual options

CapabilityTesseract AcademySemantic web consultancyGraph database vendor
Correctness of the finished modelClosed-world vocabulary gate over SHACL, measured on real corporaSHACL alone, which passes unknown termsLoad succeeds, so the data is assumed valid
Evidence you can check51 published studies, 29 open repositories, reproducible from public dataClient references, work usually under NDAProduct benchmarks on vendor-chosen data
Standards defectsReported upstream to IATA, GLEIF, W3C, GSA and USDAStandards taken as givenOut of scope
Language model integrationFine-tuned and constrained against the published vocabulary, with measured conformanceIncreasingly offered, rarely measuredRetrieval over the graph, no term-level check
DeploymentSingle self-hosted binary, no JVM, data stays in your networkVaries by engagementLicensed platform, often hosted
Public sector procurementCCS RM6200, RM6094, RM6126, RM6219 and RM6235VariesSoftware licence, not a service route

What we build

Domain ontologies

OWL 2 class and property models with a SKOS controlled vocabulary and SHACL shapes for the constraints that have to hold. Reified assertions where a fact needs a source and a date rather than being asserted flat.

Standards crosswalks

Mappings between two or more published standards, with the loss measured rather than assumed. Our waste study quantified it in bits: of 5.907 bits of composition detail, the Simpler Recycling channel retains 38.4 per cent.

Knowledge graphs

Instance graphs built from the whole source, with provenance on every assertion. Our space object graph carries 833,403 triples over all 70,122 catalogued objects in Earth orbit.

Verification for AI pipelines

A closed-world gate that rejects ontology terms your model invented. This is the layer that stops a plausible, fluent, wrong triple from entering the graph, and it is the part almost no RAG or agent stack currently has.

Register and reference data assurance

Auditing the registers your systems join against. We have measured Companies House, the Global LEI System, the FDIC, EIOPA, ISTAT, CelesTrak and others, and found identifier failures in all of them.

Training and capability transfer

A 15-lesson ontology engineering course covering RDF, RDFS, OWL 2, SPARQL, SHACL, triple stores, GraphRAG and governance, plus in-house workshops for teams taking over a model we built with them.

Proof

National Digital Twin Programme, Department for Business and Trade

The AI ontology extension generator, open-sourced under Apache 2.0

NDTP needed to accelerate ontology development across UK infrastructure. We contributed to an open-source tool that automates ontology generation and extension through a web interface, combining data profiling, named entity recognition and language models to extract ontology entities from CSV, JSON and RDF/Turtle. It shipped as a production Streamlit application with a four-step wizard, built-in validation and iterative refinement, published on GitHub under the National-Digital-Twin organisation.

3

Input formats: CSV, JSON, RDF/Turtle

Apache 2.0

Code licence, OGL v3.0 for docs

NDTP

Maintained by the programme

Read the case study

IES4, the UK Information Exchange Standard for defence and national security

Term confabulation from 93.7 per cent to 1.0 per cent

An off-the-shelf model asked to write IES4 invents most of the vocabulary it uses. We fine-tuned Qwen3-Coder-30B on the standard and measured against the published ontology rather than against a judge model. Confabulated terms fell from 93.7 per cent of outputs to 1.0 per cent, and term conformance rose from zero to 88.6 per cent. The checkpoint is published openly on Hugging Face, so the numbers can be reproduced rather than taken on trust.

93.7% to 1.0%

Confabulated ontology terms

0% to 88.6%

Term conformance

Open

Checkpoint published on Hugging Face

Read the study

IATA ONE Record, the air cargo data standard

Every property lost its domain axiom, and nobody had noticed

We audited six releases of the ONE Record data model ontology with two independent engines, rdflib and our own, which agree exactly. All 496 properties in the 2022-12 release declare an rdfs:domain. None of the 522 properties in 2023-12 do, and none of the 534 in 2024-12 do. Classes with no label rose from 1 to 58 over the same span and total lint issues rose from 2 to 650. The audit is reproducible against the MIT-licensed public repository, and the finding went upstream.

522 to 0

Properties declaring rdfs:domain

2 to 650

Lint issues across six releases

Two engines

rdflib and Open Ontologies, in agreement

Read the study

All 51 ontology studies

Register integrity, standards crosswalks, verification research, language model grounding and catalogue conformance, each with a full write-up and, where one exists, the repository.

Browse the ontology work

How to commission this service

  1. 1

    CCS RM6200, Artificial Intelligence DPS

    The main route for ontology design, knowledge graph delivery and verification work inside an AI programme.

  2. 2

    CCS RM6126, Research and Insights DPS

    For standards audits, crosswalk studies and register assurance commissioned as research rather than as build work.

  3. 3

    Direct award below £10,000

    A scoped audit of an existing ontology or knowledge graph, or a crosswalk feasibility check, sits comfortably below threshold.

Send us the ontology, the standard or the schema you are working against and we will tell you what it can and cannot verify, at no cost, within five working days. Contact fabio@thetesseractacademy.com.