Clinical AI, designed for medical control
Two readings.
One signed report.
Dyolis is a clinical AI platform designed to keep clinicians in control. It combines two independent AI readings, built-in safeguards and a mandatory clinician sign-off, and is designed to run entirely on the hospital's own infrastructure.
Development status. Dyolis is under development. It is not a CE-marked medical device and is not available for clinical use.
Dyolis Core
Five stages. The clinician is one of them.
Every case moves through the same engine, whatever the examination. Each stage follows explicit rules, and nothing leaves Dyolis without a signature.
Gate
An intake check that declines cases outside its supported scope, and says why it declined.
Draft
A first model writes the draft report from the examination.
Review
A second model, from a different model family, reviews it independently, finding by finding.
Sign
The clinician reviews, corrects and signs. Every result is a draft until they do.
Ledger
Which case, which model and version, which reviewer, which decision — recorded.
Where the readings disagree
Disagreement is shown, never smoothed away.
Two models that always agree would tell you nothing. The value is in the cases where they differ — so Dyolis surfaces the disagreement to the clinician instead of averaging it into a single confident answer.
Each finding links to the region it refers to. Abstract illustration — not a clinical image.
Draft report · findings
Module: Dyolis X-ray · ChestShown to the clinician with both positions and the region each reading refers to. Dyolis does not pick a winner.
A case the module does not support is refused with a stated reason, rather than answered anyway.
No report is issued from this screen until it is signed.
Illustration of the interface concept. Not a clinical case, and not a product screenshot.
How it is built
Five commitments, each written as a design, not a promise.
We describe what the architecture does. Where something is not yet true of the product as delivered, we say so.
Clinicians in command
Every result is a draft until a clinician reviews and signs it. Clinicians can edit any part of it, and when the two AI readings disagree, the disagreement is shown to them.
Stage · SignSafeguards by design
A multi-agent system with defined stages: an intake check that declines cases outside its supported scope, a first reading, an independent second reading by a model from a different family, and clinician sign-off.
Stages · Gate, Draft, ReviewDesigned for data sovereignty
Designed to be installed on the hospital's own servers. In this configuration, patient data is processed on site and is not sent to external AI services. Models run locally and the system can operate without an internet connection.
On-site deploymentExplainable by design
Every finding comes with its evidence: the image region it refers to and the second reading's verdict. Every step the AI took is recorded and can be reviewed.
Stages · Draft, Review, LedgerTraceable and secure
Every step is logged: which case, which model and version, which reviewer, which decision. Patient identifiers are removed before AI processing, following the DICOM de-identification standard. Role-based access and encryption.
Stage · LedgerWho it is for
Three audiences, one system of record.
Clinicians
Two independent readings, each finding with its evidence. You keep the final word.
- Edit any part of the draft before signing
- Disagreements surfaced, not resolved for you
- Every finding linked to its region
IT, security & data protection
Runs on your infrastructure. Every step is logged.
- Designed to operate without an internet connection
- Identifiers removed before AI processing
- Role-based access, encryption, full audit trail
- Software bill of materials for the components we ship
Hospital leadership
Faster reporting, with clinical control and data kept on site.
- One engine across examinations, one supplier
- Designed to support human-oversight requirements
- Decisions traceable for review and audit
The platform
One engine. A module for each examination.
A module is named after what it reads. The examination sets the pipeline — the file format, the models, the intake rules — so modules follow the examination rather than the organ. Chest X-ray is the first pack in development.
Modules beyond the first are planned, not available. Dyolis Case sits above the examination modules: it does not re-read an image, it works from the report a clinician has already signed, and its second reading checks that every sentence is traceable to a document in the record.
Before a pilot
The security review comes first. We build for it.
Your IT, security and data-protection teams will ask their questions long before a clinical conversation starts. These are the artifacts we produce for that review, as part of the build rather than afterwards.
What we ship with the build
- A CycloneDX software bill of materials for the services and the web application
- A deployment check that verifies the system makes no outbound calls to external AI services in the on-site configuration
- Containerised on-site deployment; models run locally
- Architecture and security documentation, with decision records
How patient data is handled
- Patient identifiers removed before AI processing, following the DICOM de-identification standard
- Designed so that patient data is processed on site and not sent to external AI services
- Designed to operate without an internet connection
- Role-based access and encryption
What can be reconstructed afterwards
- Which case was read, and by which model and version
- What each of the two readings concluded, finding by finding
- Which cases were declined at intake, and why
- Who reviewed, what they changed, and who signed
Regulatory status. Dyolis is under development, is not a CE-marked medical device, and is not available for clinical use. We are glad to share the current regulatory assessment and verification material with a prospective design partner under an NDA.
Talk to us about a design partnership.
We are looking for clinicians and radiology teams willing to review the reporting workflow and tell us where it is wrong. No clinical deployment, no patient data.
hello@dyolis.health