Surgical Tray Auditor
On-device · works in airplane mode

Count the tray.
Then count it again.

Photograph the laid-out tray before the incision. The phone names every item and counts it. Photograph it again at close. Anything missing, anything extra, is on the screen in three seconds.

A second pair of eyes. The manual count and the WHO checklist stay exactly as they are.

Run the demo→What it can’t do yet
Before · capture 01detecting
forceps ×2
scalpel ×3
cotton wool ×4
unrecognised #1
Lane's tissue forceps2
Scalpel handles3
Cotton wool balls4
Unrecognised items1

Four taps, start to close

The before photo is the expected inventory. No fixed sets, no pre-registered trays — the app works with whatever is laid out.

01
Photograph the tray
Before the incision, one shot of the laid-out tray on the mat.
02
Confirm the list
Names and counts, editable. The nurse's correction is the manifest.
03
Photograph again
At close, same mat, same shot. Runs entirely on the handset.
04
Reconcile
Before against after, item by item, including the unnamed ones.

Three answers, no ambiguity

Unrecognised objects are tracked as individual items by appearance, not ignored. An unnamed clamp that goes missing still trips the alarm.

All accounted for
18 / 18
Every manifest item found in the after photo, at the right count.
Items missing
17 / 18
Named, with the crop from the before photo. One cotton wool ball out of four counts.
Not in the manifest
18 + 1
Something on the tray that was not there at the start. Also worth knowing.
Stated plainly

What this is not

Not validated in theatre
The model is trained on desk captures of real instruments and common medical items, not on data from an operating theatre. Different lighting, drapes, blood, fluid, sterile gloves in frame — transfer is unvalidated. It is the first thing Phase 2 fixes.
Not tested against a workflow
This is a technology demonstration of the counting and reconciliation mechanism. Nothing here has been tested against a clinical workflow.
Not a counting system
Retained surgical items are a never-event with regulatory weight. The manual count and the WHO checklist stay exactly as they are. The app's job is to catch the case where the manual count was already wrong.