Voice AI in production for Australian clinics

Client

Clinara (my company)

Role

Co-founder, technical lead

Timeline

2025 – now

Stack

Deepgram (AU), AWS Sydney, Cognito, Next.js

The problem

A clinician's day is split between the patient in the room and the notes about them. The notes are part of the job. They're the clinical record, and they're what the next clinician reads. But they're written after the consultation, from memory, often at the end of a long day.

Clinara listens to the consultation, writes the note, and gives the clinician feedback on how the consultation went. The clinician reviews the note and signs it off. The model does the typing. The clinician stays responsible for the record.

As of September 2026, 150 clinicians across 30 Australian clinics use it, and it has processed more than 2,000 consultations in production.

Why onshore mattered

A recording of a consultation is health information, about as sensitive as personal data gets. Under the Privacy Act, sending it overseas doesn't send the responsibility with it: the clinic stays accountable for what happens to it offshore.

Clinics don't want to take that on, and they shouldn't have to. So the rule from day one was that patient audio, transcripts and notes stay in Australia. The application runs in AWS Sydney, sign-in runs on Cognito in the same region, and speech-to-text runs on Deepgram's Australian deployment.

That rule removed options. Plenty of speech and language services are cheaper or newer offshore. Each one had to wait until it could run onshore, or be left out.

Consent under surveillance devices law

Recording a conversation in Australia is governed by state and territory surveillance devices laws, and they don't agree. In some states a party to a conversation can record it. In others, NSW among them, you generally need the consent of everyone in it.

A product used in clinics across several states can't depend on which state the consultation happens in. So Clinara follows the strictest version everywhere: the patient agrees before recording starts, and that agreement is kept with the consultation record.

Consent is part of the product, not a policy document. If the step can be skipped, it will be, on a busy day, by someone who means well.

Evals and QA

A demo shows that the model can write a good note once. Production needs to show that it writes a good note on every consultation, including the noisy ones, the ones with accents, and the ones where the patient and clinician talk over each other.

So the notes are tested against real consultation types, not just the examples that look good in a demo, and every change to a prompt or model is checked against that set before it ships. The clinician's sign-off is the last check. What clinicians change before they sign is the most useful feedback the system gets.

Cutting speech-to-text cost without leaving Australia

Speech-to-text is the largest variable cost in a product like this, because every minute of every consultation goes through it. The easy saving was to send audio to a cheaper provider offshore. That would have broken the rule the product is built on.

So the savings had to come from inside the onshore setup: choosing the right model for each job, being deliberate about what audio gets transcribed, and working with the provider's Australian deployment rather than around it. The cost came down, and the data didn't move.

What's next

Clinara isn't ISO 27001 certified yet. The security work is heading there, so clinics can check the controls against a standard rather than taking them on trust. And the coaching side of the product is where the most is still to learn. Writing the note saves the clinician time. Helping them get better at consultations is the part only this kind of product can do.