Clinical AI that runs
where the data lives.
Health data is most sensitive at the moment it is created: a voice, a heartbeat, a night of sleep. Our platform is built on one architectural conviction. The model goes to the data, and the data never goes to the model.
Raw signals never leave the device.
The acoustic pipeline from our published study is the foundation of the Psi-Vox application we co-develop with clinical research partners, and that application is built to run end to end on the patient's own smartphone: capture, speaker separation, spectral feature extraction and neural network inference.
What crosses the network is a narrow, deliberate slice: derived features and scores, with quality indicators and provenance. Raw audio never leaves the device, and any transient speech processing is consumed and discarded on the spot.
- Speaker diarisation and mel-spectrogram feature extraction, on device
- Deep learning inference on phone-class hardware, with no cloud dependency
- Offline-tolerant by design: capture and scoring continue without connectivity
- When signal quality is insufficient, the system declines to score rather than guess
Each person is their own baseline.
Voices, heart rhythms and sleep differ enormously between people. Anatomy, accent, habit and medication all leave their mark. A model that compares people against a population average spends most of its capacity learning who someone is, not how they are.
We take the opposite path: measure each person against their own reference, and detect meaningful change from that baseline. It is the harder engineering problem, but it matches how clinicians actually think about their patients.
- Personal reference profiles instead of population thresholds
- Change-from-self as the guiding design principle across modalities
- Designed for longitudinal use: the same measure, repeated over time
We evaluate the way we'd want to be audited.
The health-AI literature is full of inflated numbers from models that quietly memorise the speaker, the site or the microphone instead of the clinical signal. We treat that as the central engineering risk, not a footnote.
Our published study used strict patient-level cross-validation. Everything we build now is held to standards that go further: no person appears on both sides of a split, confounds are hunted deliberately, and results are reported with calibration and uncertainty rather than headline accuracy alone.
- Patient-level, speaker-disjoint splits, always
- Leakage controls: device, site and identity cannot masquerade as signal
- Calibration and uncertainty reported alongside discrimination
- Pre-specified analyses over post-hoc cherry-picking
Decision support, never decision-making. Our models describe a current state for a clinician to consider. They never predict individual outcomes, never trigger automated actions, and never replace clinical judgement.
The other half of the story is hardware.
Models are only as good as the signals that feed them, and the silicon they run on.
Silicon & Sensing →