How our AI works —
Total transparency.

AI at MDMC is an assistant, not a decision-maker. Here is exactly how, where and why.

Four models, all European or open-source.

No closed proprietary model in production. No call to OpenAI or Anthropic at patient runtime.

Mistral 7B Instruct (FR)

Generation · Patient chat

Medical document summaries, contextual answers in patient chat.

OVH AI Labs · France · HDS v2
Open-source · Apache 2.0

CamemBERT-bio

French medical NER

Extraction of medications, exams, diagnoses from prescriptions and scanned reports. Fine-tuned on Quaero / E3C.

Self-hosted · OVH GPU · France
Open-source · Apache 2.0

Phi-3 mini · Qwen 2.5

Intent classification · Suggestions

Chatbot intent routing, lightweight CPU-friendly suggestions. Low latency.

Self-hosted · CPU · France
Open-source · MIT / Apache

Mistral Large (EU API)

Doctor decision support V2.5+

Clinical decision support with CE IIa marking (in progress). Always with human validation.

Mistral AI · Paris · EU
Commercial · EU contract

Measured in production, not in demo.

Metrics monitored continuously, drift detected, models re-evaluated quarterly.

Medical NER accuracy

85%

F1-score on Quaero 2024

Chat p95 latency

< 2 s

Mistral 7B streaming response

Alert false positives

< 15%

Internal target · monthly review

AI uptime

99.5%

SLO 24/7 OVH HDS v2

Indicative figures — pilot phase. Public reporting from Q4 2026 (AI Act high-risk register).

Our AI can be wrong. We say so, we measure it.

Always validate with a clinician

No AI suggestion replaces a healthcare professional's opinion. Summaries and alerts are indicative.

Possible biases

Models trained mostly on French adult data. Possible degraded performance on under-represented populations (pediatrics, regional languages, rare diseases).

Models evolve

You can disable any AI feature from Settings → AI & Privacy. 1 tap, immediate effect, no questions asked.

Four rights, actionable in 1 tap.

Article 22 GDPR

No automated decision alone producing legal effects. Systematic human validation for any impactful clinical suggestion.

You enable or disable each AI feature separately (summary, alerts, chat, NER). No bundled imposition.

Explainability

Every AI suggestion shows its source (HAS guideline, patient document, indexed article) and confidence level (low / medium / high).

1-tap opt-out

Settings → AI → Disable all. Immediate effect, past suggestions kept read-only, right to be forgotten respected.

Four frameworks, all documented.

Technical files available on request for partners, authorities and auditors.

Our philosophy goes further

Four non-negotiable principles, acknowledged limits and a contractual manifesto. AI is a tool. You remain the decision-maker.