Our own LLM · No OpenAI · No Anthropic

Our LLM. Not OpenAI. Not Anthropic. Yours.

A language model trained specifically for dentistry, running on your clinic's local GPU. Zero cloud calls. Zero per-token cost. Zero third-party dependency.

Why we don't use GPT-4 or Claude

Cloud LLMs are brilliant for generic tasks. For a dental clinic with sensitive health data and high AI query volume, they are the wrong choice.

Unpredictable per-token cost

Every query to OpenAI or Anthropic is billed by tokens. In a clinic with 50-100 AI queries a day, the cost grows significantly and, above all, varies month to month. Local hardware pays back in 12-18 months with zero marginal cost.

Vendor lock-in

If OpenAI changes prices, deprecates a model, or goes down, your clinic stops working. Your own LLM on your server doesn't depend on the health of a third party.

Health data outside the EU

Queries sent to OpenAI travel to US infrastructure. That forces SCC contractual clauses and creates GDPR friction. On-premise, data never leaves the clinic's server.

Network latency

A cloud call requires a round trip over the internet. Local is tens of milliseconds instead of seconds, and doesn't depend on your connection quality nor the provider's service status.

Trained on real dental data

Our LLM is not a generalist model with a dental prompt. It's trained on a profession-specific corpus.

18,000+
Indexed PubMed papers
Peer-reviewed scientific literature, dental and craniofacial, indexed with our own knowledge graph for contextual retrieval.
113,000+
Validated clinical cases
Real anonymized cases with their diagnosis, treatment and outcome verified by licensed dentists.
252,000+
DICOM radiographs processed
Real production volume used both to train the Vision AI and to fine-tune the LLM with radiological descriptions.

How we evaluate it

We don't publish marketing benchmarks. We measure AI-Doctor concordance in real production and publish the figures.

  • AI-Doctor concordance measured continuously with integrated outcome tracking. Every validation or correction by the dentist feeds the improvement cycle.
  • Every LLM answer includes the bibliographic sources it used. The dentist can inspect and verify them.
  • MUTEX between Vision AI and Text AI: two models never run in parallel on the same GPU. This guarantees precision over throughput, a deliberate decision.

Tech stack

Base model
Custom LLM trained on dental corpus + RAG over knowledge graph
Execution
Dedicated on-premise GPU (Linux / Windows), MUTEX with Vision AI
Average latency
1.8 s — 95th percentile below 3 s
Privacy
100% on-premise. Data never leaves the clinic's server.
Hardware partner
PCSpecialist (NVIDIA Inception program)

Frequently asked questions

Can I switch the LLM if a better one comes out?+
Yes. The architecture is designed so the underlying model is replaceable. If we release a better version in the future or you want to use another open-source model with the same fine-tuning, the upgrade happens without migrating data.
What happens if the LLM's answer is wrong?+
Every answer is reviewable by the dentist, who can accept, correct or ignore it. Every validation or correction feeds the outcome tracking used to measure and improve the model.
Is it really cheaper than paying OpenAI?+
It depends on usage volume. For a clinic with 50+ AI queries a day, local hardware pays back in 12-18 months. For multi-site clinics with thousands of queries a month, savings are significant from month one.
Is it aligned with the European AI Act?+
Yes. AI systems in healthcare are classified as high-risk by the AI Act. Having the model and the data on-premise greatly simplifies traceability, records and the conformity assessment the regulation requires.
Can the LLM cite fabricated sources (hallucinations)?+
The system uses strict RAG: it only cites papers and cases that are actually indexed in the knowledge graph. If the evidence doesn't exist, the agent says so explicitly instead of fabricating a citation.

Our LLM. Not OpenAI. Not Anthropic. Yours.