Skip to content

Geometric / Security

Security

Data handling, hosting and security contacts for our inference service.

Inference data

We operate zero data retention for inference. Prompts and outputs are never written to disk or logs, never used for model training or evaluation, and never viewed by our staff. The only exception is short-lived caching in memory, described below.

Our privacy policy covers inputs, outputs, operational metadata and our policy against using inference content for model training or evaluation.

Our inference service uses in-memory caching. Cached state can remain after a response.

Research, training and reinforcement-learning projects require separately agreed data scope and terms.

Hosting and data locations

EU inference
Verda infrastructure, Finland
US inference
Together, United States
Platform infrastructure
AWS, Frankfurt, Germany
Monitoring
Grafana Cloud, United Kingdom

We plan to migrate platform infrastructure to AWS European Sovereign Cloud.

Grafana Cloud is currently hosted in the UK. We plan to migrate monitoring to the EU.

US deployments are available where agreed. European inference traffic is not routed to US capacity.

Infrastructure and access

All API traffic is encrypted in transit using HTTPS (TLS 1.2 and 1.3). Traffic between our gateway and GPU servers runs over encrypted WireGuard tunnels.

Access to production systems is limited to the engineers who run the service, protected by multi-factor authentication, and logged.

Verda and Together AI have completed SOC 2 Type II audits and hold ISO 27001 certifications.

Security program

We are working toward SOC 2 Type II compliance and ISO 27001 certification. We work with Atoro as our security partner.

An independent penetration test is in progress.

Security questions

For security inquiries, contact security@geometric.so.

To report a vulnerability, follow our disclosure guidance.