Why I built this

Hospital discharge is a strange moment. Everyone is relieved, everyone is tired, and the family goes home with a new letter that looks almost the same as the last one.

Almost. A dose is halved. One medicine is missing. Another one is now twice a day instead of once. The old letter from last spring is still in the folder, and the boxes in the cabinet still follow the old instructions.

Nobody sits down and compares two medical letters line by line on the evening they come back from the hospital. So I wanted the NAS to do that one careful job.

What it does

Care Changes reads the new discharge letter and up to three older care letters from the family's folder on the NAS. It looks at every medicine as eight small fields: strength, amount, route, how often, how long, start or stop, condition and notes.

Then it compares old and new, field by field. For example:

  • Metoprolol: 50 mg twice daily before, 25 mg once daily now, with the exact sentence and page from both letters.

Every difference becomes a plain question on a printable page, with space for the pharmacist's answer. If something can not be read, it says "manual check required" instead of guessing.

It does not give medical advice and it does not decide which instruction is right. It only shows what the letters say, quote for quote.

How it works

I used the official UGREEN Skill Developer Package (uliya-skill-kit). The Skill is one SKILL.md file that sends each job to a NAS expert: file-agent finds the letters, document-agent reads and compares them, photo-agent can read medicine box labels, and knowledge-agent saves the ledger and the question page. Everything stays on the NAS with local models. I used Claude for research and drafting help, and Gemini and ChatGPT to criticize the idea.

The hard parts

  • Doses are dangerous to reword, so every number is copied exactly as printed. 2.5 mg must never become 25 mg.
  • "Twice daily" and "every 12 hours" sound the same, but the Skill shows them as different wording and lets a person decide.
  • Medicine box photos are useful, but they prove nothing, so box text only creates questions and never changes the ledger.

What I'm proud of

  • The Skill passes the official UGREEN checker with zero errors and zero warnings.
  • Every field has its own quote and page number, so a pharmacist can check it in seconds.
  • The ledger refuses to look complete when it is not. One unreadable field and the header says "Incomplete".

What I learned

I learned how UGOS Skills hand work to different experts, and that in medical topics the most helpful thing an AI can do is to be careful and say "ask the pharmacist".

What's next for Care Changes

If I become a Finalist, I will test Care Changes on a real UGREEN NAS with fictional letters that hide 25 planted differences, including 2.5 mg against 25 mg in a small table cell, and count every field as right, missed, wrong or invented.

Built With

  • local-ai
  • ugreen-ai-nas
  • uliya-skill-kit
Share this project:

Updates

Submission history