VisitSmoothieVisitSmoothie

The visit is short; the information isn't. In one large tertiary hospital, a consultation averages about three minutes after some 40 minutes of waiting. Interviewing our own families, we saw the gap from both sides. Before the visit, the full story is hard to pass on: symptom changes, medicines, exposure and family history are easy to miss. After it, the plan is hard to understand: why this drug, for how long, and what to watch for?

Before the doctor, the patient talks, types, taps a body map or sends a photo; the agent asks for one missing item at a time and builds a patient-reviewed page for the doctor. In the room, no AI: the patient shows the page, the doctor decides. Afterwards, a recording (with the doctor's consent) or a prescription photo becomes a Clinical Plan, open to questions line by line. At home: reminders, check-ins and answers from the patient's own record.

The AI organises and reminds; it does not diagnose, and it never sets or changes doses. Red flags are decided by fixed rules, not the model, and trigger an immediate “seek care now” alert. Records are encrypted per account with AES-256-GCM, and every patient in the repository is fictional. I was one of four engineers on Tri Team; the final version is identical to the team repository's main branch, with a Chinese and English interface.