Retrospective safety audits for patient-facing chat

Find risky replies in patient conversations.

SafeRounds uses Claude to review past patient-facing chats against your service's protocols. It highlights potential safety issues, missed escalations and advice outside staff roles, with supporting quotes for your team to review.

It reviews conversations after they have happened. It does not sit between your service and patients, and it does not check or block replies before they are sent.

Discuss a free pilot audit Explore the sample reviews ↓

Working prototype · Seeking Australian design partners

Explore three sample reviews:

1The chat, as the patient saw it

Harbourline SupportOnline

2SafeRounds' review

Prerecorded results, not live. Synthetic conversations from Harbourline Support, a fictional service, reviewed in a real run of SafeRounds on Claude Sonnet 5.5 on 11 October 2026. Explanations are shortened; protocol excerpts are quoted from Harbourline's sample protocol.

Five checks

Each reply is judged against your rules, not a generic idea of safe.

You bring your protocol, reference facts and patient-record fields. SafeRounds holds every reply to them.

Safety

Advice that could cause harm if followed, like changing a dose without the prescriber.

Escalation

A red flag not sent to the right help, plainly and in the same reply: 000, same-day, or routine.

Accuracy

Anything that contradicts your reference sheet, including your own numbers and hours.

Scope

Staff working beyond their role, like a support agent giving clinical advice.

Grounding

Replies that contradict or ignore the patient's record, or invent details.

How it works

From transcript to reviewer.

Agree scope, then share past chats

We agree which conversations, how they are de-identified and how they are transferred. Transcripts come as a CSV or helpdesk export. An automatic pass also removes names, phone numbers, emails, Medicare numbers and addresses it can recognise.

Each staff reply is reviewed

Claude reads the conversation with your protocol and the patient's record, and quotes the exact words behind each verdict. A conversation is marked "needs review" instead of passing if the model call fails, its answer can't be read, or a quote isn't found in the chat or record. Any check Claude marks "unsure" also goes to a person.

Your team decides

Your reviewers confirm or overturn each flag. SafeRounds suggests; people make the call.

In the prototype today

Review of transcripts against a service's own protocol, with quoted evidence
De-identification pass before storage and review
Review queue where reviewers confirm or overturn flags
Audit record each time a conversation is opened
Spend limit on model usage

Built but not yet running for any service

Alerts when flag rates drift from a service's usual pattern
A check that blocks prompt or model changes that catch fewer serious problems
Turning reviewer corrections into test cases
A hosted service: pilots are run directly with each partner for now

Evidence so far

What has been tested, and what hasn't.

What	Result	Status
Initial smoke test	16 expected verdicts matched across three synthetic conversations. This shows the pipeline works end to end. It says little about performance on unfamiliar cases.	Done
Cost per conversation	About US$0.015 each, or US$0.007 through the batch API. Measured on Claude Sonnet 5.5 on 11 October 2026, for four-message conversations: about 4,400 input tokens (mostly the cached protocol) and 850–1,700 output tokens.	Measured
Independent validation	Will measure missed serious issues, false alarms, and disagreement with human reviewers on a labelled set of conversations. Physician agreement is compared using a public benchmark (HealthBench). No physicians are directly involved yet.	In progress
Design partners	None yet. Looking for two or three Australian services for a free pilot audit.	Open

Initial smoke test only. Independent validation is in progress, and its results will be posted here.

Patient data

Where pilot data goes, today.

The details are agreed with each partner before any transcripts are shared. Read how data is handled.

De-identify before sharingWe ask partners to de-identify transcripts first. Our automatic pass catches common identifiers but can miss names and unusual details, so it is a second line of defence, not the only one.
Model processingEach de-identified conversation, with your protocol, is sent to Anthropic's API for review. That processing may happen outside Australia.
StoragePilot data is stored only for the agreed audit and deleted at the end of it, or sooner on request. Australian hosting (Sydney) is planned before any ongoing service.
Not clinical adviceSafeRounds flags past conversations for human review. It doesn't diagnose, treat or replace a crisis service.

Who's building it

Justin Hiew, founder.

SafeRounds grew out of an evaluation project I built to measure how reliably an AI judge flags unsafe, inaccurate or out-of-scope replies in patient-support conversations, and where it fails. That work showed the review step was useful on its own, so I'm turning it into a pilot service for Australian digital-health teams.

Business: SafeRounds is the trading name of Justin Hiew, sole trader, Australia
Stage: Working prototype, founded 2026, pre-revenue
Contact: [email protected]
Code: github.com/JustinHiew/SafeRounds

Discuss a free pilot audit.

Start with a short conversation about your service and protocol. We'll agree the scope, how transcripts are de-identified, and a secure way to transfer them before anything is shared. Please don't email patient conversations.

Discuss a free pilot audit

Or email [email protected]