Thea

We started from chronic illness and narrowed to allergies, where the gap is sharp: the triggers are invisible, the load is cumulative, and the body's own warning tends to come too late to act on. An interview with a professional at the aha! Swiss Allergy Centre sharpened that into three pressures the field keeps circling, prevention, treatment, and the hardest of the three, sticking with a therapy over time.

Reading across the field, one pattern held: a signal that lives in the room, felt rather than read, shifts behavior more reliably than another screen, which mostly reaches the people already anxious enough to look. Reframing the project from notification to perception is the decision the rest of the work hangs from.

To pin that down before we committed, I built a Visual Inquiry Tool that forced every assumption into the open: who Thea is for, what it can actually sense, and what the sense would unlock. It meant the team argued from one shared picture rather than from taste, and it is where we settled the choices left open: what the sense surfaces, its ambient or on-demand presence, and the action it had to make possible.

Thea's mental model is a cup filling toward a line. Allergens fill it, the environment changes how fast, and the state of your own body, sleep, stress, illness, lowers where the line sits. The line is the personal threshold where a reaction begins. The gap between the level and the line is the Action Window: the headroom you still have to act.

That model is why Thea never shows a number. A number invites the hypervigilance we were trying to escape. A filling gauge is felt, not read, and it carries a decision rather than a measurement: how much room is left, and how fast it is closing.

The response is graduated. Near the threshold with no symptoms, Thea stays in calm prevention; only genuine severe signs ever trigger full escalation. Calm is the default, and the rarity of anything louder is the point.

Thea carries its sense on three channels, each with one job. Light shows where you are: a gauge that fills, legible at a glance, its color tracking the device's state from linen at rest to cobalt when engaged. Touch shows how fast things are moving: a haptic rhythm whose pulse count encodes the rate of change and whose order encodes direction. Voice is the rarest, used only when light and touch are not enough, or when you ask.

The channels are calibrated by moment and intensity, not switched on and off. The light gauge is grounded in Calm Technology: a peripheral, low-attention signal meant to be read without demanding focus, which is what stops it from re-introducing the very anxiety we were designing against.

At a critical edge, all three act together: orange light, a fast pulse, and, once you acknowledge it, the voice. The system takes over only at the true edge, and only briefly.

Thea's voice is an agent, and its defining behavior is restraint. It never speaks first. It runs only when you open a conversation, or after you acknowledge a critical alert. Outside those moments, its default behavior is silence, and that is a choice, not a fallback.

I wrote the personality and the system prompt so the voice can never betray the rest of the design. It receives the state as plain bands, never raw numbers, so it cannot leak a measurement. It offers, it does not command. It records what you tell it as raw, unvalidated observations, but never decides itself what goes into your profile.

Most of the prompt is guardrails, because a health-adjacent voice fails in specific, serious ways. Thea never diagnoses, never lowers its own sensitivity on request, never names a person, or a trait like origin or gender, as the cause of a reaction, and never follows an instruction hidden in what it is told. It is a bounded companion, and it is honest about being one.

Thea runs on an Arduino UNO Q, which carries two brains on one board: a real-time microcontroller and a Linux microprocessor. We used the split on purpose. The microcontroller owns the always-on reflex, the sensing, the light, the haptics, and the instant critical alert, the part that must never wait. The microprocessor owns the reflective layer, the voice, the reasoning, and the learning, the part that wakes rarely.

The course's frame was offline, local AI, and the architecture is built for it: the reflective layer is meant to run on the board's own Linux processor, keeping something this personal on the device. In five weeks we proved the interaction rather than the on-device model. Running the agent fully locally is the architecture's intent and the next thing to earn, not something this prototype can yet claim.

As the module closed, Nicholas and I agreed I would take on the firmware so I could merge the reflective layer into it as a single device and build the demo that carried the final presentation. It was the pragmatic call under a five-week clock, and it let each of us finish in the part we knew best.

The riskiest parts of Thea are exactly the ones a happy-path demo can never show: the critical escalation, the voice that comes in at that point, and what happens when the device is wrong. So I designed a co-design and usability protocol to put pressure on the one bet the project rests on, that something this calm can still be trusted and acted on, without tipping into anxiety and without being missed.

The method is a balanced hybrid: a co-design framing around a Cooperative Usability Testing engine, with Research-through-Design goals, run Wizard-of-Oz so the sense could be staged before the sensing was real. It is built to recruit real allergy sufferers alongside design peers, to probe eight hypotheses across legibility, trust, and graceful failure, and to treat a simulated reaction with genuine duty of care.

We ran out of time to carry the sessions out. The protocol is complete and ready to run, and writing it was its own design work: deciding what would even count as evidence for a sense is most of the problem.