Why emergency AI must work offline

I still ride interventions. Last winter, a resuscitation in an underground parking level, two floors down. No signal. None. That is not an edge case in emergency medicine; it is a Tuesday.

Every clinical AI product I had tested until then assumed a connection. Dictation to the cloud, transcription in the cloud, structuring in the cloud. Two floors underground, all of it becomes a very expensive voice recorder that will sync "later". Later is after the handover, which is exactly when the information mattered.

The documentation tax

Emergency clinicians pay a documentation tax on every intervention. Vitals, timings, medications, scores, decisions: all captured by memory and reconstructed afterwards, often 40 minutes after the patient left the vehicle. Reconstruction under fatigue is where data quality dies, and where the legal exposure lives.

Ambient AI is the obvious answer. Listen, structure, compute, report. Hospitals are deploying it in consultation rooms today. But the emergency chain is a harder environment in three specific ways:

1. Connectivity is adversarial. Basements, tunnels, rural dead zones, mass-casualty saturation: denied environments are the operating baseline. The network is absent exactly when acuity peaks.
2. Consent and privacy are non-negotiable. Streaming a patient's worst moment to a third-party cloud, from a public street, is a GDPR problem you do not want to explain to a data protection authority.
3. Latency is clinical. A score that arrives after the handover is trivia, not decision support.

What on-device changes

SONARA runs its full pipeline (speech recognition, clinical extraction, score computation, report generation) on the device in the responder's hand. This is not a degraded offline mode; it is the architecture.

• No network requirement, full performance in any field condition
• Patient data stays on the device until the clinician decides to share it
• Sub-second clinical feedback during the intervention, not after it
• Four languages (FR, NL, DE, EN), because Belgian reality is multilingual

The trade-off is real: on-device models must be small, fast and disciplined. We spend our engineering budget on that constraint instead of on GPU clusters. It is the harder path and the only one that survives contact with an underground parking level.

Where we are

Field validation with Belgian emergency teams is underway. The device rides along, listens, and we measure everything: transcript quality in sirens and wind, extraction accuracy against the official report, time returned to the clinician.

That data, not demos, will decide what SONARA becomes. And the moment all of it converges on is the handover.