"We work on data our clients have not published yet. Adding AI could not mean adding someone else who can read it."
Eugeniu Costetchi, Founder & CEO, Meaningfy
Partner Story
How Meaningfy runs AI on client data without adding a vendor who can read it.
Meaningfy is the Luxembourg semantic engineering house behind interoperability systems used across the European public sector, from the eProcurement Ontology to the TED Semantic Web Services. It runs its AI workloads on Tresor's confidential inference, so the documents, models and records its clients hand over stay readable to Meaningfy alone, with verification receipts that show exactly where and how each request was processed.
Snapshot
Meaningfy
- Industry
- Semantic technology, data interoperability and knowledge engineering
- Region
- Luxembourg / EU - serving EU institutions, public administrations and private clients
- Tresor product
- Confidential Inference API
- Use case
- Knowledge extraction, document classification, LLM-assisted semantic mapping and semantic search over confidential client material
- Outcomes
- Confidential AI workloads on Tresor's zero-access inference, with no third party in the path who can read client content
- One confidential foundation reusable across every client engagement, rather than a separate approval per project
- Verification receipts that answer the sub-processor question with evidence instead of a policy
The Challenge
The data that makes AI useful here is data Meaningfy does not own.
Meaningfy turns complex, fragmented data into knowledge graphs, ontologies and semantic search systems. That work happens inside other organisations' material: the source documents, the internal records, the models still under development.
For EU institutional clients that means regulatory and procurement material ahead of publication, document corpora that exist to be harmonised rather than exposed, and reference models that are still being negotiated between stakeholders. For private clients it means the business data being lifted into an enterprise knowledge graph in the first place, which is usually the data the organisation guards most closely.
Meaningfy is not the owner of any of it. It is the party trusted to handle it. That distinction is the whole problem: a product company adding AI answers to its own board, while a consultancy adding AI answers to every client whose data would pass through the new system.
The same data that makes model assistance useful here is the data clients most expect to remain contained. If AI becomes another readable vendor in the chain, the work no longer sits inside the trust boundary Meaningfy was hired to respect.
What Meaningfy works inside
- Source documents and record corpora held by public bodies
- Regulatory and procurement material ahead of publication
- Reference models and ontologies still under development
- Client business data being modelled into knowledge graphs
- Personal and sensitive data encountered during content classification
- Research datasets held under grant and data-use agreements
Why that matters
Meaningfy's clients did not hand over their data so that it could be forwarded somewhere else. Any AI in the workflow has to sit inside the trust boundary that already existed, not beside it.
Why Public AI Was A Non-Starter
Both of the obvious answers failed for the same reason.
Meaningfy could either route client material through a public model API or absorb the full stack itself. Neither option kept the client trust boundary where the engagement began.
Public model APIs
Routing client material through a public AI endpoint puts a third party who can read every word into a chain that was, until that moment, just Meaningfy and its client. For an EU institutional client, that is a new sub-processor: a contract amendment, a transfer assessment, a fresh approval, and often a straightforward no. Repeated across a portfolio of engagements, it stops being paperwork and starts being a reason not to use AI at all.
Running models in-house
Owning GPUs, operations, model updates and incident response would turn a small team of semantic engineers into an infrastructure team. It would also leave the client with nothing but Meaningfy's word that the data stayed put.
There is a particular trap in this work. Meaningfy's classification services inspect content precisely in order to find the personal, sensitive and confidential parts of it. The model has to read exactly the material that must not leak. There is no version of that workload where the confidentiality question can be deferred.
The Solution
Confidential inference inside the existing trust boundary.
Meaningfy runs its AI workloads on Tresor's Confidential Inference API. Because the API is OpenAI-compatible, it went in as a configuration change rather than a rebuild.
Zero-access processing
Every prompt and document is processed inside an attested hardware enclave. Tresor cannot read the content, and neither can the inference provider behind it. No new party gains sight of client material, so the trust boundary stays exactly where the client agreed it would be.
Verification receipts
Each request produces a signed, independently verifiable receipt showing which model ran, where, and inside which attested environment. For a team whose clients are auditors, institutions and public administrations, evidence carries a conversation that assurances do not.
EU processing and open models
Inference runs in EU regions on an open-source model stack. For a company that builds European data infrastructure and publishes its own tooling as open source, that is consistency rather than a compliance checkbox - and Tresor's move to Luxembourg-operated compute puts the AI in the same jurisdiction as the work.
How It Came Together
2 lines
of configuration
to route existing AI calls through confidential inference
Meaningfy's engineers changed a base URL and an API key. Nothing else about their stack had to move, because the confidential foundation went underneath the existing workloads rather than replacing them.
- No rebuild around a new provider surface
- One confidential path reusable across client engagements
- Confidential Workspace under evaluation for internal team use
What's Now Possible
AI across the practice, on material that could never have left.
Knowledge extraction, mapping, search and classification can now run across the same confidential client material Meaningfy already handles under contract.
Knowledge extraction on closed corpora
Entity recognition, information extraction and semantic annotation can run over document sets that were previously out of reach for anything but on-premise tooling.
LLM-assisted mapping and modelling
Mapping data to ontologies is interpretive work. Model assistance can now be applied to client models and source formats that have not been published, inside Mapping Workbench and the surrounding toolchain.
Semantic search over client knowledge graphs
Retrieval and question answering can run against the graph itself, with the confidential content staying confidential through every inference call.
Classification of sensitive content
Content-based classification can identify personal and sensitive material without that material becoming visible to an AI vendor along the way.
The Quieter Win
For a consultancy, "which AI are you using?" is not a curiosity. It arrives in every security questionnaire and every procurement annex, and until recently the honest answer created work for everyone.
Meaningfy now answers with an architecture and a receipt. The confidential foundation was established once and applies to every engagement after it, which turns a per-project negotiation into a property of how the company works. For a team whose stated belief is that meaning and responsibility are two sides of the same coin, that is not a procurement convenience. It is the position matching the principle.