Laboratory professional working with AI-assisted digital systems

Artificial intelligence for regulated healthcare

We install AI models on your infrastructure or on EU servers and integrate them into your processes with traceability, human oversight and regulatory compliance.

Let's discuss your AI use case See an example

Two services, one principle

Health data stays where it belongs, and every AI output can be explained, reviewed and audited.

Local AI: installation, configuration and training

We deploy open-weight models on your servers or on dedicated servers in the EU. Data never leaves your perimeter and you do not depend on an external provider or per-use licences.

We adapt the model to your work: we connect it to your documentation (RAG) and, where it adds value, train it on examples validated by your team.

Model selection by task and language
Server sizing (GPU, memory)
Inference server with an internal API
Search across your documents (RAG)
Fine-tuning on your own data
Evaluation, validation and versioning

Advanced capabilities via API

We add AI to the applications you already use through APIs: calling your local models or existing models from providers that host and process data in the EU.

Between the application and the model sits a control layer that decides what data is sent, logs every request and ensures a person validates before the result takes effect.

Sensitive information never goes to external models. The AI suggests and the person decides: no model output is applied without validation when it affects clinical or quality data.

Integration into existing applications
External models only for non-sensitive data
Pseudonymisation before data is sent
Logging of input, output, model and user
Human validation of every decision
Access control and encryption

Information control

Sensitive information is never shared with external AI models

Pasting patient data, results or internal documents into a public AI tool means handing them to a third party. In healthcare and regulated environments, that's not an option.

The risk of external AI services

Data leaves your organisation and may be stored on servers outside the EU.
Depending on the service and licence, the provider may keep conversations or use them to improve its models.
You lose control over who can access the information and for how long.
It may breach the GDPR, professional secrecy or confidentiality agreements.
There is no internal record of what was sent, who sent it or which answer was used.

Our approach: control and local models

Sensitive information is processed by local models, on your servers or dedicated EU servers: it never leaves your perimeter.
Today's local models handle tasks such as classifying, extracting data, coding, summarising and searching documents very well.
If a specific case needs an external model, it only receives non-sensitive or pseudonymised data, processed in the EU, with a contractual guarantee that it is not used for training.
A control layer filters what goes out, blocks what must not leave and logs every request.
A clear usage policy tells your team what they can and cannot do with AI.

Information we always process locally

Patient data
Clinical and laboratory results
Staff personal data
Formulas and manufacturing processes
Quality documentation
Prices, contracts and bids

Local models let you benefit from AI without giving up confidentiality: the same kind of help, with your data kept at home and predictable costs, with no per-use fees.

How we install a local AI

A well-scoped project, with deliverables at each stage and documentation built for audits.

Use case and risks
Task, data involved, risk level and impact assessment.
Model choice
Size, language, performance and a licence suitable for professional use.
Infrastructure
Your own server or a dedicated EU server, on an isolated network.
Deployment
Inference server, internal API, SSO and logging.
Adaptation
RAG over your documents and training on validated data.
Validation and monitoring
Testing, metrics, periodic review and change control.

What you need for a local AI

The components we size and document in every installation.

Hardware

A server with a GPU sized to the model. For small models or classification tasks, a CPU may be enough.

Model

An open-weight model with a licence suitable for professional use, chosen by task and language.

Inference server

The software that runs the model and exposes it as an internal API, with queues and concurrency control.

Data

Good-quality documents and examples with clear permissions, for RAG or training. Validated by your experts.

Security

Isolated network, SSO, encryption in transit and at rest, and a log of every request.

Governance

Owners, usage procedures, acceptance criteria and change control for every model version.

Regulatory compliance by design

EU servers

Data processed and stored in the EU. With local AI, it never leaves your infrastructure.

Full traceability

Every suggestion is logged with input, model, version, output, user and final decision.

Human oversight

Professionals see the reasoning behind each suggestion and make the call.

GDPR and health data

Data minimisation, pseudonymisation and impact assessments where required.

EU AI Act

Risk classification, transparency and technical documentation for the system.

GxP and validation

Risk-based validation and model change control, aligned with EU GMP Annex 11.

Example

AI-assisted LOINC coding of laboratory tests

Assigning the right LOINC code to every test in a catalogue is expert, repetitive work. The AI suggests; the professional decides.

Load the catalogue
The laboratory provides each test’s local name, specimen, method and result units.
Reasoned suggestion
The model searches the LOINC table and suggests the best-fitting codes, explaining which axes match: component, property, system, scale and method.
Human validation
The user reviews the information shown and validates, corrects or rejects. Nothing is coded without their confirmation.
Log and improve
Every decision is logged with user, date, model and version. Validated mappings improve future suggestions.

LOINC mappingLaboratory catalogue › GLU-S

Test: Glucose
Specimen: Serum
Method: Enzymatic (hexokinase)
Units: mg/dL

Model suggestions

2345-7Confidence 92 %
Glucose [Mass/volume] in Serum or Plasma
Component, system (serum or plasma) and property match: mass/volume, consistent with mg/dL.
2339-0Confidence 61 %
Glucose [Mass/volume] in Blood
Units are consistent, but the system is whole blood and the specimen is serum.
14749-6Confidence 23 %
Glucose [Moles/volume] in Serum or Plasma
Correct system, but the property is molar (mmol/L) and does not fit mg/dL.

Validate 2345-7Choose another code

Log: suggestion from local model v1.3, awaiting user validation.

Sample screen with fictitious data.

LOINC® is a registered trademark of Regenstrief Institute, Inc. The same pattern works for SNOMED CT, ICD-10 or internal catalogues.

More examples

Other processes where the AI suggests and the user decides

The same pattern with different data: the AI does the searching and comparing, and the expert validates with all the information in front of them.

Laboratories · Hospital networks

Unifying laboratory test codes across several systems

After a merger or across a network of sites, each laboratory system (LIS) uses its own codes and names for the same test. Without a common catalogue, results cannot be integrated or the data analysed.

Starting data: Each system’s catalogue: local code, name, specimen, method, units and reference ranges.
What the AI suggests: Groups equivalent tests across systems, proposes a master catalogue with its LOINC code and flags conflicts: different units (mg/dL versus mmol/L), or specimens and methods that do not match.
What the user decides: Confirms or splits each group, resolves conflicts and sets the unit conversions.
Outcome: A versioned master catalogue and equivalence table, ready for integration, the electronic health record and data analysis.

Purchasing · Sales

Cost estimation and quote generation

Preparing a quote or a bid means cross-checking purchase history against market prices. Done by hand, it takes hours and depends on who prepares it.

Starting data: Purchasing records and history, prices awarded in public tenders and competitor prices from public sources (published catalogues and price lists).
What the AI suggests: Matches equivalent products even when their names differ, estimates cost and a market price range, and drafts a quote with the target margin, citing the source of every figure.
What the user decides: Adjusts items, margins and terms, and approves the quote. Nothing is sent without their approval.
Outcome: Faster, more consistent quotes, with every price justified and a decision history to refine future bids.

Have a process AI could speed up?

Coding, document classification, searching procedures or reviewing records: we'll assess your case and tell you whether it makes sense and how to do it safely.

Let's discuss your case