Always watching the sources you choose
Gmail, Google Calendar and course deadline feeds, each connected with its own permission. Dira polls them in the background and keeps a commitment graph of what you promised, to whom, and what depends on what.
A professor moves your midterm two days earlier. Suddenly your interview is too close, your study time is gone, and a teammate is waiting on work you can’t do. Dira notices, finds every promise the change breaks, and repairs your week within limits you set. It tells the people affected before they have to ask.
The demo is a synthetic student week with a simulated calendar and mailbox. It needs no account and cannot contact real people.
Gmail, Google Calendar and course deadline feeds, each connected with its own permission. Dira polls them in the background and keeps a commitment graph of what you promised, to whom, and what depends on what.
NVIDIA Nemotron models on Nebius Token Factory read the change and propose repair plans. Deterministic code checks every plan for feasibility, provenance and policy, and you approve anything sensitive. No model output can authorize an action.
Reject or edit a step and Dira suggests a rule. Confirm it and it is committed to a private Git repository you can export, and mirrored as a read-only Skill for your Hermes assistant.
Engine proofs, including the deterministic replay, the 20-run reliability check and the variation matrix, are on the Evidence page.