Common Ground

Inspiration

A shared plan often breaks in one small place. Someone can still do the pickup but cannot make the return trip. Restarting the whole conversation wastes the agreements people already made, while explaining a refusal can reveal more than they wanted to share. Common Ground explores a more careful way to adapt.

What it does

The demo follows a fictional care circle. Ravi can take Dev to an appointment but cannot bring him home. Leena offers the return on one condition: Asha must receive Dev by 16:15. Her private rule stays in her owner space; only the reviewed offer is released.

NVIDIA Nemotron on Nebius proposes a revised arrangement. Owners authorize exact commitments and handoffs, while a separate checker tests the proposal against released offers. Asha must review the changed direct counterpart even though the handoff time is unchanged.

The accepted revision inserts Leena's return and deletes Ravi's old return. Ravi's pickup retains the same event identity and version, including across restart. These are actual entries in separate local SQLite calendars. Readback receipts distinguish requested changes from observed results.

How I built it

A shared broker and separate owner services keep private memory, released offers, decisions and calendar effects distinct. React and TypeScript provide the interface; Python services use SQLite, signed decision records and a structured feasibility checker. The isolated runtime uses separate containers, mounts and owner networks under one host administrator.

Nemotron Ultra on Nebius Token Factory completed the conditional repair. The first invalid proposal was rejected and remains in the record. A local structured solver also works without inference; no model advantage over that baseline is claimed.

Challenges and what I learned

A responsibility can keep its time but acquire a different dependency, requiring a fresh decision. A model proposal, an owner's consent, a calendar entry and a person actually arriving are different facts. Keeping those distinctions visible was the central design problem.

I am proud of the selective change and its portable record. Sanitized evidence can be checked offline without private keys or the original database. The saved arrangement is now historical because its signed authority expired; opening it cannot authorize fresh changes.

What comes next

All people and decisions shown are fictional. There is no household study, live Google Calendar connection or clinical benefit claim. Next steps are voluntary household testing, independent security review and qualification of real calendar integration. The goal is a useful shared plan without turning private reasons into shared data.

Built With

• fastapi
• nebius
• nvidia-nemotron
• ortools
• python
• react
• sqlite
• typescript