AI that runs on your own servers.

Three packages. Private AI assistant for $3,000 puts an open model on your own server with a team chat. Private knowledge base for $8,000 adds your documents with permissions, sources and single sign-on. Enterprise private AI for $20,000 serves several departments, with GPU sizing, an API and audit logs. You pay 50% when you approve the plan and the rest at delivery. Hardware is bought in your name; we size it and tell you the cost per token before you order.

Open-weight models: Llama, Mistral, Qwen, Gemma, DeepSeek and smaller task models for embeddings, reranking and classification. We pick per use case on eval scores, latency and the hardware you have. A 7B to 14B model on one GPU covers most document and support tasks; a 70B model needs a multi-GPU node. Every model call goes through one gateway, so swapping a model later is a configuration change, not a rebuild.

It depends on the model and the number of people using it at once. A single server with one or two data-center GPUs serves a 14B model to a few hundred users. A 70B model wants four to eight GPUs. If you already have GPU capacity, in a private cloud or on a workstation, we start there. The research report includes a sizing table with cost per token, so you can compare it with the API bill you pay today.

It removes the hardest part: no personal data leaves your infrastructure, so there is no third-party processor, no transfer outside your country and no provider training on your prompts. We add what auditors ask for: access rules tied to your SSO, a log of every prompt and answer, retention policies and a data map in the documentation. Your legal team still owns the policy; we make the technical answer to every question a yes.

On general knowledge, the frontier APIs are ahead. On your documents, your terms and your language, a local model with good retrieval and a small fine-tune usually matches or beats them, because the answer comes from your data, not from the model's memory. We do not ask you to trust that: before launch we run evals on a set of your real questions and show you the scores side by side.

Start with retrieval. RAG puts the right documents in front of the model and gives you citations, and it covers most use cases on its own. Fine-tuning pays when the model must learn a style, a format or a domain vocabulary, such as German contract language or Turkish accounting terms. We propose it only when the evals show a clear gain, and we run it on your hardware with your examples.

Your IT team can, with the runbook we hand over: how to update a model, roll back, add a GPU, read the monitoring and restore from backup. Many clients keep us on a monthly plan instead: we keep models, drivers and the serving stack current, re-run the evals when your data changes and watch GPU use and latency. Either way, the system is yours and runs without us.

The first meeting is free. Within a week you have a research report with the use cases, the data map, the hardware sizing and a fixed price. You approve the plan and pay 50%. A first use case is live on your hardware in two to four weeks; a full platform with several use cases, SSO and fine-tuning takes up to three months. The remaining 50% is due at delivery, after the evals pass the thresholds we agreed.