AI that keeps patient data inside your hospital.
Clinicians spend hours on documentation. ARK Core runs transcription and language models on the hospital's own infrastructure, so patient data stays inside the building.
In a survey of 226 healthcare executives, AI returned 3.5 times its cost on average, mostly from administrative work rather than clinical decisions. Documentation is where hospitals start.
Bain & Company with Bessemer Venture Partners, The 2026 Healthcare AI ROI Scorecard, 30 September 2026 (self-reported).
of audio transcribed by Whisper large-v3-turbo, in 99 languages, with timestamps.
Read moreprompt fields in request logs by default. Generated text is not logged.
With a second copy of each part of the model, a failed server doesn't take the service down. Requests in flight are replayed on another path, and servers join or leave without a restart.
Read moreWhat stands in the way.
Special-category data
Health data is among the most protected data under GDPR. It can't be sent to an outside model provider.
Clinical time
Documentation competes with time for patients.
Medical-device rules
Software that informs diagnosis or treatment can fall under the MDR. Under the EU AI Act, obligations for AI in regulated products such as medical devices apply from 2 August 2028.
EU AI Act, Annex I; dates under Regulation (EU) 2026/1744.
Health data rules are changing
The European Health Data Space sets new rules for how patient data is shared and reused across the EU. Patient summaries, ePrescriptions and most secondary use apply from March 2029.
Regulation (EU) 2025/327 (EHDS), in force since March 2025.
What hospitals build on ARK.
Each runs on the hospital's own GPUs. Clinicians review every output.
Dictation and transcription
Whisper transcribes dictation and consultations on the hospital's own GPUs.
Note summarisation
Long patient histories become discharge-letter drafts for clinicians to review. A 64,000-token history reaches its first token in 4.7 seconds.
Coding support
Classification codes are suggested from notes for staff to confirm.
Internal knowledge
Staff questions are answered from the hospital's own procedures, using retrieval over bge-m3 embeddings.
What your data protection officer will see.
Inside the hospital
With ARK Core, the gateway, compute servers and models all run on the hospital's infrastructure and are operated by your team. Nothing depends on outside services.
Logs without content
Every request carries a correlation id. Prompt fields are redacted by default, and generated text is not logged.
Known, open models
Every model is open-weight, with its licence and model card published. The model that answered is named in every response.
Evidence before go-live
ARK's API test suite runs on your deployment and writes standard JUnit results.
Built to stay up
ARK splits each model into stretches of layers spread across the hospital's GPUs and machines, and can hold a second copy of any stretch. If a server fails mid-request, the request is replayed on another path and the service keeps running. Servers can join or leave a running deployment without a restart.
The rules healthcare AI works under.
ARK is an inference platform, not a certified medical device. Uses that inform diagnosis or treatment need their own conformity assessment. Your compliance assessment stays yours.
Two ways to run ARK. One API.
ARK Core For hospitals
The platform and its models on the hospital's NVIDIA GPUs, licensed per GPU and run by your team. Our engineers can size it with you, prove it on your hardware and support the rollout.
ARK CoreOur EU infrastructureARK Cloud
Prototype on the hosted API with synthetic or de-identified data, then move in-house without code changes.
ARK CloudStart with documentation.
Bring a sample of de-identified dictation or notes. We run it on ARK, on your GPUs or ours, and show you the transcripts, the drafts and the numbers before you commit.