National AI: use your own country's model, under your own rules
Delentia does not tie you to one model or one country. A Thai organisation can choose a Thai model, a Japanese one a Japanese model, or run a model on its own machine. The safety gate stays the same because it sits outside the model. That is neutrality you can check in the code.
Built and tested
- The model is a setting, not a dependency. Choose a model on your own machine (Ollama), any model on OpenRouter, or any OpenAI-compatible endpoint, which is how a national or self-hosted AI plugs in. On the Models page you pick country, then provider, then model, or enter your own.
- The safety rules do not depend on the model. The safety gate, the human signatures and the audit log are code outside the model, so they apply the same way whichever model is chosen. The model changes how capable the agent is, not how safe it is.
- A data-residency policy per organisation, checked before every model call: home region, allowed regions, whether any cross-border call is allowed, and what to do with personal data (allow, redact or block). A blocked call is never sent. Each decision is written to the audit log without the personal data itself.
- A personal-data scanner with national checksums (Thai ID, Japanese My Number, Chinese resident ID, Korean resident number), plus payment cards, e-mail addresses and phone numbers.
- A router that suggests a model by detected language and region from a table you can edit, and backup models that may never expose data more than the main model does.
What it does not do
- It does not know any country's law. It enforces the rules your organisation writes. It is not legal advice, and you decide which rules apply to you.
- It cannot see what a provider does with data once it arrives. That is a matter for your contract with the provider.
- It covers the model calls the agent makes. Other network tools are not model calls and are not covered by this policy.
- No national model is bundled and none is a partner. The default routing table is the author's starting guess, and no quality comparison per language has been run.
- Results were tuned and tested on a few models, mostly small local ones. Report any result per model, never pooled.