Private AI
OrcStat's AI plans analyses, checks assumptions and explains results using a model on your own computer or network. Your data stays where it is, and a validated engine computes every number.
Models
One click downloads a llama.cpp runtime with Qwen2.5 (1.5B or 7B) or Phi-4-mini, sized for your hardware. No other software to install.
Found automatically when they run on your computer. Use any model you already have.
Point OrcStat at vLLM, llama.cpp or any OpenAI-compatible server in your lab, hospital or office.
What the AI does
Ask a research question in plain language. The orchestrator profiles your variables, proposes an analysis plan, checks normality, equal variances and outliers, switches to the right alternative when an assumption fails, and drafts an APA 7 results section.
You approve each step. Dialogs and approved AI actions are recorded as syntax in the journal, so the whole analysis can be re-run without the AI. The model writes the words; the engine computes the numbers, and every number the AI writes is checked against the output.
Controls
On by default. The AI may reach only model servers on your machine or approved private networks; DNS answers and redirects are re-checked, and blocked attempts are logged.
Each project decides how much a model may see, from variable names and labels up to full data. The default is names, labels and output tables, never your rows.
See exactly what a request contains before it's sent. Choose what the AI may do without asking: Ask, Balanced or Hands-off.
Every AI request records what was shared and what you approved. Export it for an ethics board or purge it.
IT can lock the model endpoint, keep AI on the network, cap exposure and switch off AI features for every install.
OrcStat works with no model: test selection and checks fall back to built-in rules, and every statistic still comes from the engine.
Licence checks send only the licence ID, a hashed device ID, the app version and the operating system. See security and data handling.
By default it sees variable names, labels and output tables only. You can raise or lower that per project, and preview every request before it's sent.
OrcStat scores each model on statistics tasks and recommends a size for your hardware. An 8B-class model runs well on a laptop with 16 GB of RAM; smaller models still help with explanations.
Only if you choose to. Cloud models from OpenAI, Anthropic or Google are off by default, need your own API key, are turned on per project, and can be forbidden by an admin policy.
The AI never calculates. A validated engine computes every statistic, and each number in the AI's text is checked against the engine's output before you see it.
Every module and the local AI orchestrator. No card needed.