Designing near the things people protect
Handle
with care.
Health information. A pharmacy record. Someone’s location. As the work moved closer to daily life, the smallest assumptions began to matter more.
Start with what should never be collected.
Work around dental services, pharmacy workflows, scanning and location made quality feel less abstract. A label, a record or a suggestion could influence a decision beyond the screen. The design needed to be careful about what it knew and what it merely inferred.
I explored structured domain data, validation, controlled storage and assistive interfaces. Across these projects, a useful starting point was the smallest amount of information necessary to complete the task. Collecting more because storage is available creates obligations without necessarily creating value.
A boundary belongs in the data model as well as the copy. Which fields are needed? Who may retrieve them? What remains on the device? What changes when someone shares? Answering these questions early affects the architecture, not just the wording of a permission dialog.
Default state: keep the information on the device.
A choice has to change the behavior.
Location experiments were a concrete way to think about control. Local-first behavior keeps the starting point on the active device. Sharing should be a distinct action with an understandable scope, not a silent side effect of opening the application.
That distinction matters because a reassuring sentence cannot compensate for contradictory behavior. If the interface says something is temporary, the storage and deletion model need to support that statement. If a person can withdraw a choice, there must be a real path for doing so.
Convenience often pushes in the opposite direction: keep more context, remember more history, make the next action automatic. Sometimes that trade is useful. In a sensitive workflow, it should be deliberate and visible.
Say what cannot be established.
Assistance can reduce cognitive load, but polished language can also create authority the system has not earned. A specialist assistant needs to show its limits where those limits affect the next decision. Hiding uncertainty at the bottom of a page is too late.
Strict validation creates its own usability obligation. Refusing an unsafe or incomplete action should come with an explanation of what is missing and a practical next step. A safe failure still needs to be usable.
Scanning provides a useful example of the distinction between recognition and knowledge. Reading an identifier is a technical operation. Establishing what the associated record means, whether it is current, and whether it is sufficient for the next action involves other rules. A successful scan should not imply that every later question has been resolved.
The same care applies to the language of absence. “No result found” is different from “there is no relevant record.” The first describes the system’s observation; the second makes a much broader claim. Keeping the sentence close to what the system actually knows is a practical engineering habit.
These explorations changed where I look for trust. I look at defaults, retention, validation and the ability to stop. Each is a small, observable promise. Together, they say more than a friendly interface ever could on its own.
Related work & references
Selected public entry points and related portfolio pages. A company website or profile is context, not a claim that private source code is available.
Mumya Pharma and DentistAI are private projects. Public demos and source repositories are intentionally omitted.
Scope, context & further questions
Grounded in public service and interface work across healthcare-oriented, pharmacy, scanning, storage, and location experiences.
- Sensitive explorations: 4 domains. Dental, pharmacy, barcode, and location-oriented work tested different trust boundaries.
- Privacy direction: Local-first. Location behavior was designed to remain on the active device unless explicitly shared.
- AI posture: Assistive. Specialist workflows were framed as decision support rather than autonomous authority.
Questions I am still working through
Can the safest path also be the easiest path?
Often yes, if safe defaults are embedded in the interaction itself: collect less, keep sensitive state local when possible, validate before submission, and make the recommended path the low-friction path.
The difficult cases are where safety requires extra verification that cannot honestly be hidden.
How should sensitive products prove that deletion and consent controls are real?
Controls need observable system behavior: explicit retention rules, auditable deletion states, clear scopes, and interfaces that show what will be removed or retained before confirmation.
Independent verification and operational evidence remain necessary for higher-trust systems.