Your team should build use cases, not run GPU drivers
A private AI platform is a model server such as vLLM or Ollama, a vector index such as Qdrant, a chat layer, identity, pipelines, tracing and GPUs, each with its own release cycle. Someone has to patch them, back them up and answer the alerts. With Pilae, that is us: monitoring around the clock on every plan, and around-the-clock incident response on Enterprise. Your data and AI team keeps ownership of the models, the prompts and the use cases.
The private AI offer covers the organisation-wide assistant. This page is about the platform underneath it, and what you can build on top.
Models where your data is allowed to be
We run open-weight models on the hardware you choose: your own GPU servers on premises, dedicated GPU machines in Pilae Cloud in the region you choose, or both in a hybrid setup. For sites with no outbound connection, the whole stack runs air-gapped on Enterprise plans. The contract names the location, and data residency is not a setting a vendor can change.
Retrieval on internal data
Retrieval-augmented generation grounds answers in your own sources: Nextcloud, Outline, file shares and databases such as PostgreSQL. The index lives next to the data. Permissions come from your directory through Keycloak or your existing identity provider, so retrieval never widens anyone’s access.
Agents that reach real systems, on record
MCP servers give Claude, Cursor or your own agents a standard way to read and act on your apps. Model Context Protocol is an open standard, so the same servers work with any client that supports it. Each person signs in over OAuth, the agent acts with their permissions, and every tool call lands in the audit trail your security team can export to its SIEM. Pipelines in n8n and Windmill prepare the data and run the agents on a schedule, and Langfuse records each model call with its prompt, answer, cost and latency, so a new model or prompt can be compared with the one it replaces. Before an agent touches production, it can run in a sandbox: an isolated copy of the app and its data, inside your perimeter.
Operated under contract
The Pilae Agent plans every update and waits for your approval. Monitoring probes every service every 60 seconds, and alerts reach Pilae engineers around the clock. Backups are daily, encrypted, offsite in the country you choose, and restore-tested every month. To scope a pilot, tell us about your project.