Sovereign AI, on every cluster.
Soverstack deploys a sovereign LLM and an ops agent on your own GPUs - automatically, with the rest of the cluster. No API key, no exfiltration, no SaaS dependency.
What apply deploys
Three workloads. Zero manual setup.
When you run soverstack apply, the AI stack is rolled out alongside the cluster. Defined in workloads/regional/, just like any other service.
- Sovereign LLM
workloads/regional/llm.yamlVM with gpu-large flavor + 200 GB disk. Hosts Llama, Mistral, Qwen, or any open-weights model you choose. OpenAI-compatible API exposed on the mesh. - Ops agent (HA pair)
workloads/regional/agent.yamlLeader + standby VMs with profiles you can toggle: ops, observability, security, compliance, performance, backup, customer. - GPU flavor
flavors.gpu-largePCIe passthrough configured at Proxmox level. NVIDIA / AMD / Intel - whichever you put in the rack. - Mesh integration
tailscale / headscaleAI reaches your database, secrets, monitoring and storage over the encrypted mesh. No public Internet hops.
Use cases
What teams build with AI on board
Internal RAG
Search and synthesize across corporate docs, code repos, tickets - without a single byte leaving your perimeter.
Pair programming
OpenAI-compatible endpoint plugs into IDE tools (Continue, Aider, Cursor self-hosted). Your code stays sovereign.
Customer support AI
Tier-1 triage and reply drafting trained on your knowledge base. Customer data never reaches a third party.
Ops automation
The agent watches your metrics, suggests scaling decisions, drafts runbooks, and writes post-incident reports.
Anomaly detection
Time-series and log analysis on top of your SIEM. Catch the unusual before alerts fire.
Compliance audit
Continuous checks against ISO 27001, RGPD, HDS, HIPAA controls. Audit-ready evidence on demand.
Sovereignty guarantees
Three things that never change.
Your data never leaves
Inference, training, fine-tuning - all on hardware you own. No telemetry, no remote inference, no opaque vendor backend.
Your model, your choice
Pick Llama, Mistral, Qwen, Gemma, or your own fine-tune. Swap it anytime by editing workloads/regional/llm.yaml.
Compliance by design
RGPD, HDS, HIPAA-ready by default thanks to LUKS-encrypted storage and isolated mesh networking.
Run AI you actually own.
Get a live walkthrough of the LLM and agent in a real cluster.