EU private cloud
- Hosted in EU regions, dedicated GPU inference
- Managed updates, MLOps and corpus refresh
- 99.5% availability target
- Data residency chosen per tenant
Your infrastructure. Your rules.
Sensitive data should not travel to someone else's cloud to be analysed. Draven Cyber's platform is being built to run where the client decides, under European law, with security and AI governance designed in from the first line of code.
On-premises appliance
- Validated GPU server image inside your building
- Works fully offline, with secure offline updates
- External model APIs disabled by default
- Built for institutions that must keep data in-house
The numbers we build to
| Area | Target | Detail |
|---|---|---|
| Groundedness | ≥ 90% | Claims verifiably supported on the evaluation set; unsupported claims never emitted silently |
| Full draft assessment | ≤ 10 min | First complete, sourced draft for a standard dossier |
| Interactive response | ≤ 5 s p95 | Chat-style workflow steps |
| Concurrency | 200 users | Per inference node |
| Encryption | TLS 1.3 · AES-256 | In transit and at rest, with managed secrets |
| Architecture | Zero-trust | SBOM, SAST/DAST in CI, penetration testing before every major release |
| Standards | ISO 27001 · NIS2 | Certification readiness and NIS2-aligned incident processes |
| Privacy | GDPR by design | No training on client data, processing agreements, DPIA templates |
| AI governance | EU AI Act | Technical documentation, logging, human oversight and transparency notices to ship with the product |
Open weights, our adaptation
We are fine-tuning licence-compatible open-weight models on a Romanian administrative and legal corpus using parameter-efficient methods, and training our own smaller models for verification, classification and ranking. Serving is designed to be model-agnostic, so no single vendor can switch us off.
Mistral / MixtralAPACHE 2.0
Qwen familyAPACHE 2.0
Llama 3.xCOMMUNITY LICENCE
Romanian-adapted modelsOPENLLM-RO / ROLLAMA
In-house verifierPROPRIETARY · NLI
LoRA / QLoRAADAPTATION METHOD