Newest Ordicura
Personal AI Assistant
A per-user assistant reachable from a browser, a terminal or a chat client, deployed on infrastructure the business owns.
In detail
-
Routed, not fixed
Each request is classified into one of four tiers — simple, normal, difficult, private — and answered by the model designated for it. The designation is configuration, changed at any time and effective immediately.
-
Single-tenant per user
Data, credentials, budget and retention are isolated per user. No shared store spans them, so a fault in one instance reaches no other.
-
Confirmation before side effects
Reads execute automatically. Anything that contacts a third party, incurs expenditure or mutates a host raises a confirmation and waits. What has been allowed to run unattended is itself a list, and can be withdrawn.
-
Scheduled and conditional work
A prompt plus a schedule runs unattended under the ordinary permissions and confirmations. The same construct with a condition instead of a clock watches for a change rather than a time.
-
Provenance-tracked memory
Every retained assertion records its origin and can be traced back to it. Ingested content is never promoted to an assertion, so a document cannot write itself into what the assistant believes.
-
A ledger of what it did
One record per action that changed something: what was attempted, whether it succeeded, and what can be reversed.
What it looks like
