HealthTech
We helped MediMind, a Norwegian health-tech startup co-owned by practising doctors, build its own AI capability: models fine-tuned on the founders' own expertise, running on a zero-trust platform where patient data never leaves Norway.
- Client
- MediMind AS
- Duration
- Long-term, ongoing
- Result
- GDPR · NIS2 · HIPAA
Services
Domains
Tooling
- Kubernetes
- Zero-trust
- HSM (FIPS 140-2 L3)
- SLSA 3
- Self-hosted models
- Fine-tuning
Result
- ( r1 ) MediMind runs its own AI, on its own hardware, inside Norwegian jurisdiction. No patient data leaves the country, and no external API sits in the path.
- ( r2 ) Small fine-tuned models match large general-purpose ones on the task, at a fraction of the cost and latency, on infrastructure the client controls.
- ( r3 ) Every expert correction now compounds into training data automatically, and no new model ships until it clears an expert-calibrated, held-out evaluation gate.
- ( r4 ) A zero-trust platform built past GDPR, NIS2 and HIPAA: FIPS 140-2 Level 3 key protection, kernel-level threat termination, a SLSA Level 3 supply chain, and a unified audit trail that meets the 72-hour breach-notification window.
- ( r5 ) Founded and co-owned by practising doctors, MediMind can show an enterprise buyer or a regulator exactly how its AI is trained, run, and secured, with a path to FedRAMP open.
MediMind is a Norwegian health-tech startup, founded and co-owned by a group of practising doctors who put their own money and judgement into it. They needed AI good enough that the doctor-owners themselves would trust it, working on patient data that, by law, cannot leave the country. General-purpose AI rented from a large foreign provider fails both tests. So we helped them build their own, and the platform to run it on.
In European healthcare, data sovereignty is a legal requirement, not a preference. A provider headquartered abroad can be compelled to hand over data regardless of which region the server sits in, so 'available in the EU' is not a solution. That ruled out sending patient data to a general-purpose model API, and pointed to one answer: models MediMind runs itself, on infrastructure it controls.
The people who own MediMind are the people who would judge its output every day. That sets a high bar. Good enough for a demo is not good enough for a doctor who has put their name on the company. Quality had to be built against their standard and measured against it, not against a generic leaderboard.
In a regulated domain, a quality score means little without a trustworthy, repeatable way to measure it, and a platform you can prove is secure. Both the models and the ground they run on had to hold up to an auditor, not just a pitch.
Two things had to be true at once: models good enough for the people who own the company, and a platform secure enough to run them on patient data. We built both, and MediMind owns both.
We moved MediMind off rented, general-purpose APIs and onto small open-weight models they run themselves. Fine-tuned on the founders' own expertise, a small model reaches the quality of a large general-purpose one on their narrow task, at a fraction of the cost and the latency, and every request stays on hardware they control. Owning the model is what makes the sovereignty real: there is no external API anywhere in the path.
The real asset is not the model, it is the loop that keeps improving it. We built the tooling where a domain expert reviews the model's output and corrects it, and every correction becomes a labelled training example. That data feeds the next round of fine-tuning, which produces better output, which needs fewer corrections. The people who know the field generate the training data as they work, automatically. A competitor renting an API cannot copy that, because they never see the corrections.
New models do not ship because they look better in a demo. We built MediMind an evaluation harness they own: expert-calibrated, severity-weighted scoring against a held-out set the model has never seen, with the automated judge itself validated against expert labels before it is trusted. A candidate model runs silently against live traffic in shadow mode, and is only promoted once it clears the gate. It survives every model swap, and it is what keeps quality honest. One early lesson made the point: a flattering benchmark was quietly misleading us until we fixed the measurement itself, then re-based every number on the corrected setup.
- held-out set
- severity-weighted
- judge validated
Owning the models only matters if the ground under them is trustworthy. We built MediMind a zero-trust Kubernetes platform for patient data, engineered to GDPR, NIS2 and HIPAA and, in several places, beyond them. Every service authenticates and encrypts to every other with mutual TLS. Patient data is encrypted per record, and the keys live in a FIPS 140-2 Level 3 hardware security module, the standard used by banks and defence. Nothing trusts the network it sits on, and access is denied by default everywhere.
Most systems detect a threat and page a human. This one terminates a malicious process at the kernel level in microseconds, before any data leaves the machine. There are no long-lived passwords anywhere: workload credentials expire in an hour, deployment tokens in ten minutes, so a stolen secret is dead before it can be used. And every release is cryptographically signed and independently verifiable to SLSA Level 3, so the cluster refuses to run any software it cannot prove came from the pipeline.
All of it runs self-hosted inside Norwegian jurisdiction: inference, training, labelling, logs, backups. Nothing crosses the border. Seven audit streams are unified into one view, patient identifiers stripped out, and kept for a year, so an incident can be reconstructed and reported inside the 72-hour legal window, and a regulator or an enterprise customer can be handed a live view rather than a promise. It clears the security questionnaires that decide enterprise tenders, with a path to FedRAMP already open.
