Generate without drifting beyond scope.
The system needed enough flexibility to feel personal, but not enough freedom to redefine the clinical relationship.
Selected work · ReMind
AI aligned
AI aligned
with clinical
intent.
How we designed and built a voice AI system that could be personalised by clinicians, useful between sessions and deliberately constrained.
Design premise
The useful system was not the most conversational one.
Open-ended AI was the wrong interaction model for a clinician-supervised product. The work began by defining what the system must never become: a substitute therapist, an unbounded chatbot or a central store of sensitive patient history. Those constraints became the architecture.
Expose control without exposing prompts.
Clinical expertise had to shape behaviour through a usable product surface rather than technical model controls.
Coordinate care without collecting a corpus.
Useful sharing had to coexist with an architecture designed to minimise centrally held sensitive content.
Three system problems
The technical work behind one quiet interaction.
The visible experience is a patient speaking and receiving a short reflection. Making that moment dependable required three connected systems: a resilient voice application, an expert-configurable generation layer and a data architecture that limits what the platform can know.
01 · Voice system
An AI pipeline that feels like a dependable daily practice.
Problem
Voice capture, processing and generated playback cross device and network boundaries. A fragile connection or unclear state could make a sensitive interaction feel broken—or make a patient unsure where their recording had gone.
Design response
We separated the local experience from network-dependent generation. Native capture, playback and reflection history remain responsive and available offline; asynchronous processing is treated as an explicit state rather than hidden latency. Device security and biometric access are part of the foundation, not a later layer.
Interaction architectureLocal ↔ Network
01 · Local
Secure capture
02 · Async
Bounded generation
03 · Local
Playback & history
Network work is exposed as state. The daily practice remains available without it.
02 · AI control layer
Translate clinical judgement into bounded model behaviour.
Problem
Free-form prompting asks clinicians to become prompt engineers and makes output quality difficult to predict. Hard-coded templates are safer, but cannot reflect different therapeutic methods, voices or patient needs.
Design response
We built a structured authoring layer that turns clinician choices into a versioned behavioural configuration. At runtime, patient input is combined with the correct active configuration inside a constrained generation process. Clinicians control intent and boundaries; the product owns the technical translation.
Control architectureExpert-configured
VoiceHow the reflection speaks
MethodHow the reflection reasons
FocusWhat the reflection attends to
Versioned configurationApplied at runtime →
03 · Data architecture
Clinician review with minimal data retention.
Problem
The easiest architecture would centralise recordings, transcripts and reflections. It would also create a durable repository of exceptionally sensitive content—and make broad access the default.
Design response
We inverted that model. Patient content remains under patient control, while sharing is explicit and item-specific. Selected recordings are protected before transmission, readable only at the destination and automatically removed from temporary platform storage within a short window. The infrastructure coordinates access without becoming the long-term content owner.
Selective sharingPlatform cannot read
Patient
deviceEncrypts selected item
deviceEncrypts selected item
◇
Ciphertext onlyClinician
deviceDecrypts for review
deviceDecrypts for review
Temporary availabilityAutomatic expiry · 72h
System validation
Did the architecture survive contact with daily use?
A three-week feasibility and acceptability study with 27 participants tested the assembled system in practice: whether people could use it, whether the experience felt coherent and whether the interaction was worth continuing.
82System Usability Scale score
Top-quartile usability
Top-quartile usability
4.04uMARS app quality score
Out of 5.00
Out of 5.00
70%Scored 7+ for intent to continue use between sessions
These results are early feasibility signals, not claims of clinical efficacy. They support the product mechanism and provide a foundation for further evaluation.
Reusable architecture
Patterns for high-trust AI products.
The most valuable output was not one application, but a repeatable way to build AI systems where domain experts set the boundaries, users retain control of sensitive data and model behaviour must remain legible.
Voice AI systems
Domain-expert configuration
Privacy-preserving product design
Pre-regulatory research infrastructure