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 checks the commitments and dependencies you have reviewed, then proposes a feasible repair within limits you set. It prepares updates for the people affected, with drafts only by default.
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 have their own permission. Background checks require owner opt-in and deployment enrollment; they are not enabled by connecting an account. Dira keeps a graph of what you promised 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, with a reusable Skill for an optional Hermes companion. The owner runtime is a separate enrollment step.
Engine proofs, including the deterministic replay, the 20-run reliability check and the variation matrix, are on the Evidence page.