Inspiration

In the January 2025 Eaton Fire, Anthony Mitchell told 911, "There's two disabled people in the house." His son Justin needed two or three people to get out of bed. Their usual caretaker was in the hospital that night. Both men died.

The person who knows how to move someone isn't always the person who is there. I wanted a way for a family to see, on a calm day, what a stand-in helper would not know.

What it does

Stand-In is a Skill for Uliya on a UGREEN AI NAS. You ask, "What would someone need to know to get Uncle Ray out if I'm not here?"

  1. It asks who the sheet is about and which folders to look in.
  2. It reads the manuals, label photos and voice memos in those folders.
  3. It answers twelve questions a helper or dispatcher would ask, like the weight of the chair, how many people it takes to move him, his medicines and who to call.
  4. It marks each answer Written, Only spoken or Not found, with the file and page, or the recording and the time.
  5. You can fill a gap on the spot. After you approve, it saves one dated sheet to print.

The gaps are the result. The sheet shows what lives only in one caregiver's head.

How I built it

  • A Skill is Markdown only: one SKILL.md and seven reference files. No code and no outside service.
  • Uliya hands the work to four of the NAS's local experts: files, documents, photos and knowledge.
  • Text recognition and the vision model both read an equipment nameplate. A value counts only if they agree.
  • A lift check adds the chair's weight to the rider's and compares it with 600 lb, the least a vehicle lift must be built to carry under federal rules.
  • It never guesses, never converts a unit, and makes one write, only after approval.

Challenges

The Skill has to be honest about what it doesn't know. A small model playing Uliya would sometimes skip the approval step or pick a folder nobody named. I tested it against nine made-up households and tightened the rules after each round.

What I'm proud of

It passes UGREEN's official checker with 0 errors and 0 warnings, and it passed 24 of 27 simulated dry runs. In none of those 27 runs did it save anything before approval.

What I learned

Most tools report what they found. For this problem, what's missing matters more.

What's next

Nothing has run on a real NAS yet. If Stand-In reaches the finals, I'll test it with a real manual, a real nameplate photo and a real voice memo, and I want a firefighter or dispatcher to review the twelve questions.

Built With

Share this project:

Updates

Submission history