Skip to content

Evidence before deployment.

Working principles

  • Measure before building

    Agree representative data, real questions, expected outcomes, and failure cases with the people who know the domain before choosing a model or architecture.

  • Keep people in control

    Show the source, make uncertainty visible, and leave consequential decisions with the person accountable for them.

  • Make change traceable

    Version datasets, training runs, prompts, models, and retrieval settings. Keep evaluation results so behaviour can be reconstructed later.

  • Design for the real boundary

    Treat privacy, residency, network isolation, security review, and operations as inputs to the design, not launch-day surprises.

Training clinical models.

A model is only as trustworthy as its data, its labels, and the way it was tested.

  • Documented data

    Record where the data came from, how it was selected, what was excluded, and who approved its use.

  • Clinician ground truth

    Labels come from clinicians, and agreement between reviewers is measured rather than assumed.

  • Honest evaluation

    Report performance on held-out data and across patient subgroups, not as a single headline number.

  • Ethics and regulation

    We support HREC applications and help scope whether a model is a research tool, a clinical aid, or software as a medical device under TGA rules.

Australian data handling.

Sometimes the answer is not AI.

A search index, cleaner data, a simpler model, or a better workflow may solve the problem with less risk and cost. We will say so early.

Start a conversation.

Start with the questions your ethics, privacy, and governance teams will ask later.

Enquiries are typically answered within one business day.