The story

Is it time for this medicine?

Take a photo of the box. Med Check looks at your record and tells you out loud if it is time.

For people on several medicines, and those who care for them.

A demo with made-up patients and records. Not for real medicines.

Hands holding a medicine box at a kitchen table in morning light. — Same box, different time. Real app results. Made-up patient.

Sample amlodipine box — Same box, different time. Real app results. Made-up patient.

How it works

The AI only reads the name. Our rules do the rest.

Margaret, a made-up patient, at 08:00 on Tuesday.

1Photo
You take a photo of the box.
2Read
The AI picks the name from our list
- Amlodipine
- Amitriptyline
- 14 more
- None of these
3Check
Our rules check her record. No AI.
- On her record?Yes
- Due at 08:00?Yes
- Already taken?No
4Answer
NAME MATCHES
You see it and hear it. We wrote the words, not the AI.

The four answers

NAME MATCHES

It is on your record, and it is due now. It does not check the strength.

NOT NOW

It is on your record, but it is not due now. You hear when the next dose is.

NOT ON RECORD

It is not on your record. Ask your pharmacist.

CAN'T TELL

It cannot read the name or check the time. Try again, or ask your pharmacist.

Why it matters

The doctor checks it. The pharmacist checks it. At home, the last check is you, holding the box.

8.4 million

people in England got five or more medicines in three months, in late 2019.

Overprescribing Review, 2021

11 reports

of serious harm from daily use of methotrexate prescribed once a week. Some overdoses were fatal. UK, 2006 to mid-2020.

MHRA, the UK medicines regulator, 2020

Amlodipine
Amitriptyline

The names most often mixed up in pharmacy reports in England, late 2021.

HSC Northern Ireland, 2022

We found no UK count of these mistakes at home.

What we tested

50 web pictures. No wrong medicine names in this run.

27read the right name
8had a clear name, but it still said CAN'T TELL
15had no name from our list, and it said CAN'T TELL
0wrong names

A small test. 15 more pictures, kept back to the end and run on an earlier list, also gave 0 wrong. A test that small can't rule out about 1 wrong in 5.

Web pictures, not phone photos. AI wrote the expected answers, and we reviewed them. No clinician did.

Limits and privacy

It checks the name and the time, not the strength.

Strength is not checked: The right name in the wrong strength, such as 10 mg, not 5 mg, can still show a match.
A demo, not for real medicines: All patients and records are made up. A real product could well count as a medical device.
Today's photo: Today the whole photo goes to OpenAI to read the name. The app does not save it.
Pharmacy labels: A pharmacy label can show a name and address, so a real version would crop the photo first.

Our plan for a real product. Not built yet.

In a real product, your record stays on your phone.

On your phone: Your record stays here. We never hold it.
What the AI sees: Only the medicine name. Not you, and not your record.

How Codex helped

A review found a gap: a low name score could still show a match. Codex, OpenAI's coding agent, reproduced it, then added the missing check and tests. Now a low score gives CAN'T TELL.