vigilance

The clinical audit that stays on.

Patient pathways lose time and people at the handoffs. vigilance reads the records, shows a clinician where each patient stalled and why, and reports a number the trust can sign.

Open the demoSee the results
The review queue: stalled patients ranked with the reason printed on every row

Nobody has time to read every note

Clinical audit is how the NHS finds where pathways break. It is statutory, it works, and it stops the moment a ward gets busy, because the data collection is a clinician reading notes by hand.

8min 35 s

to abstract one discharge summary by hand.

28hours

per audit project, mostly unpaid time.

2020

National audit was switched off by letter when the pathways were breaking fastest.

How it works

The clinician writes the audit as a protocol: which handoffs, what target, which reasons count as legitimate, who owns the fix. Everything downstream reads from it.

1 · ReadMedGemma reads the pagesTyped, scanned or handwritten. The 4B vision model transcribes every page with word positions.
2 · PointThe model quotes, the code countsMedGemma 27B returns three verbatim quotes per case: start, end, reason. Code checks each quote on the page and computes the hours itself.
3 · ReviewOne row per patientThe clinician sees the quote highlighted on the real page, confirms or overrides in a minute, and the queue prints why each row is next.
4 · ReportA number you can signReviews correct the rate for the cases nobody checked. Alerts name an action and an owner. The report drops into the audit the trust already owes.

Why you can trust the number

Nothing is invented, nothing is hidden, and the model never has the last word.

Nothing invented

A quote the verifier cannot find on the page gets a red badge and no rectangle. A missing time is an abstain, never a guess.

Stalled, not just late

Every breach carries a reason with a class. An unsigned chart is a stalled patient. Platelets at 38 is a legitimate wait. Only the first counts as time lost.

Honest intervals

Prediction-powered inference keeps the confidence interval valid whatever fraction of rows the clinician managed. Where the model is biased, the model-only interval misses the truth every time; corrected, it holds.

It orders the team's checking, never a patient's care. This is audit, not surveillance.Aggregate outputs only. No per-patient alert, no recommendation, no patient interaction.

Read the audit report