Inspiration

Many families have a parent who lives alone, and with that comes a quiet, constant worry. The usual answer is a cloud camera, but a lot of parents refuse because they do not want a company, or anyone else, looking inside their home. Motion alerts do not help much either. They go off all day and never answer the one question families actually have, which is whether today looks like a normal day.

A NAS with cameras and local AI is the right place to answer that question. The footage already lives there, it never has to leave the house, and the analysis can run on the device itself.

What it will do

Quiet Care Guardian is designed around three modes.

  • Setup. It learns the person's normal routine from two weeks of camera events, such as when they get up, when they visit the kitchen and how long they usually stay still.
  • Rules. The family adds rules in plain language, like "let me know if she goes out after 10pm" or "she should be in the kitchen around lunch". Each rule is translated into a check the cameras can actually answer, and the family sees how it will be checked before saving it.
  • Daily check. Someone asks Uliya "how is Mom doing today?" and gets a short text report marked Normal, Worth a look or Needs attention, with the real event times behind every finding.

The video never leaves the NAS. The family gets conclusions, not footage, and if something looks off they can open that single snapshot inside the NAS app.

Where it stands

So far I have designed the full workflow and written the initial Skill package with the official uliya-skill-kit. Every step uses a capability that exists in the public catalog. Camera work is delegated to security-monitor-agent through ug-surveillance, and the care profile and rules are stored on the NAS through file-agent and ug-filesystem, with approval required for every write. The package passes the kit's static checker with no errors or warnings.

It has not run on a NAS yet. If selected as a finalist, I will use the testing day to confirm that the cameras produce the expected events and to validate the routine and rule logic against a scripted week of events.

Challenges so far

  • Presence is not action. The cameras report people by room and time, not what they are doing. Rules about eating or medicine are checked as presence in the right room at the right time, and the report says so openly instead of overpromising.
  • No scheduling or notifications in the catalog. My first idea was a guardian that warns the family on its own. Since a Skill only runs when someone asks, I redesigned it around an on-demand daily check and kept the logic ready for proactive alerts if UGOS exposes a trigger later.
  • Privacy by design. Evidence is only requested for user review inside the NAS, never for export, and only for the event the user picks.
  • Learning a routine takes time. With less than a week of history the profile is marked as low confidence, so the Skill never pretends to know more than it does.

What I learned

Designing on a platform with strict capability boundaries changes the process. Being honest about what can and cannot be detected made the idea better, because a worried family needs to trust every line of the report.

What's next

Testing on the NAS in the final round, then proactive daily summaries as soon as scheduled triggers and notifications are available.

Built With

  • file-agent
  • local-ai
  • markdown
  • privacy
  • security-monitor-agent
  • ug-filesystem
  • ug-surveillance
  • ugos
  • ugreen-ai-nas
  • uliya
  • uliya-skill-kit
Share this project:

Updates

Submission history