Building clinical AI for European hospitals: local-first and honest

The deployment reality in Europe

Many hospitals in Central and Eastern Europe, and plenty across the rest of the continent, cannot or will not push imaging data to a third-party cloud. A clinical AI that assumes a cloud round-trip is, for them, simply not deployable. We treat local, hospital-controlled inference as the default rather than an afterthought.

Local-first is a performance choice too

Running on hospital hardware is not only about privacy. Atlas processes a CT study in a median of about 2.8 seconds with TensorRT FP16 on a single consumer-grade GPU, within the memory envelope of an ordinary clinical workstation. Latency that low is compatible with sitting quietly inside a triage workflow rather than adding a queue.

Crossing borders, on purpose

A European-first product still has to generalize. We deliberately tested Atlas across countries, training in Türkiye and validating on a US cohort with frozen weights [1], because a model that only works on its home distribution is a liability. Honest external evaluation, with strict patient-level separation to avoid leakage [2][3], is the part of the work that earns a hospital's trust.

Saying what it cannot do

Trust also comes from restraint. Atlas integrates into existing structured-reporting practice rather than asking departments to rebuild around it [4], and every performance figure we publish is framed as research-stage, with thresholds and calibration treated as deployment-specific [5]. The product is a triage aid that keeps a doctor in the loop, not an autonomous reader, and we say so plainly.