Ponte
Situation-adaptive healthcare access for foreigners living in Italy. A self-directed study in what a product designer is for, in a world where anyone can use AI.
ROLE
Lead Researcher & Designer
PROJECT
Case Study
What is a designer for, when anyone can prompt?
Using AI is no longer what sets a designer apart. Anyone can prompt a tool. So I asked a harder question: in that world, what does a product designer actually need to be good at? This project is my answer, worked out in the open.
THE MESSAGE
AI does the volume. I own the decisions.
AI can compress the bulk of the work into a fraction of the time. The value I add is the last mile: the judgment about what is safe to ship, the craft that makes it real, and the preparation that lets engineers build it.
THE PROCESS
One AI-augmented pass, then the last mile by hand
AI carried the volume in the early steps. The last two are where the design actually got decided and made.
01
Research
CLAUDE
02
Wireframe
CLAUDE
03
Prototype
GOOGLE STITCH
04
Vibecode
LOVABLE
05
Critique
CLAUDE
06
Refine
FIGMA
01
Research
I used Claude to map the landscape quickly: the Italian health system and its gatekeeper GP, the medico di base, the competitors, and the gap. A few findings anchored the concept.
Italy has 5.56 million foreign residents, about 9.4% of the population, including roughly 96,000 foreign university students. The system runs on GP registration, and prescriptions from abroad are not honored.
The existing products were either broad-but-shallow booking apps or traveler services that route people around the public system. Nobody owned depth for the long-term newcomer who has to live inside it. That gap became the brief.
FRONT OF HOUSE
The input plus her situation, registered with the health service or not, produces one clear recommendation: pharmacy, GP first, specialist, or emergency, with a plain reason. Routing adapts to the situation, not just the symptom.
BACK OF HOUSE
When she picks a clinic, the same structured input is translated into the doctor's language and sent ahead of the visit. She walks in already understood.
02
Wireframe
I shaped the hero flow as six screens, following one anxious newcomer from "I'm sick and lost" to "I know where to go, why, what it costs, and my symptoms are already translated."
I set the design guardrails at this stage too, because in health the constraints are the design: structured input as the spine, an always-visible emergency exit, and an explicit consent step before any health data is shared.
03
Prototype
I prototyped in Stitch, driven by a single source-of-truth design file so every screen stayed consistent. Stitch is strong on mobile and multi-screen work and exports natively to Figma, which fit my pipeline. This is where the six screens became an interactive flow.
​
Stitch is fast and its output is genuinely attractive, but that speed comes with errors. My rule: fix the obvious problems right away, and flag the subtler ones for the refining pass, where they get the attention they need.
01 SITUATION
Country, stay, and health-system status. This is what makes routing adaptive.
02 SYMPTOM
Structured by design, so the same input can be safely translated later.
03 GUIDANCE | SIGNATURE
One urgency-colored recommendation, with the emergency exit one tap away.
04 CLINICS
Honest signals first: English level, real price or "not listed", booking method.
05 CONSENT
An explicit consent gate, off by default, listing exactly what is shared.
06 HANDOFF
A summary in Korean and Italian, side by side, plus what to bring.
04
Vibe-coding pass
To stay fluent across the tool landscape, I took the same design into Lovable, a vibe-coding tool, to see how it handles the build. The goal was not to ship from it. The import came out close but not exact, and Lovable changed some things on its own.
That is the nature of the category: a vibe-coding tool keeps its source of truth in running code, so it regenerates rather than preserves. An AI UI design tool like Stitch keeps the source of truth in an editable design file. Knowing that difference is how I decide which tool to reach for. I kept Figma as my source of truth and corrected the drift there.
05
Design critique
Then I stress-tested the prototype with Claude against the design system and the safety rules. This is where the real decisions got made.
Caught in critique
Generating sample content for the handoff screen, the AI produced a summary describing chest pain and shortness of breath radiating to the left shoulder, and the flow routed it to a GP booking. That is a textbook cardiac emergency, and routing it to a booking breaks the product's first safety rule. The machine produced something plausible and dangerous. I rewrote it to a throat-infection story that genuinely belongs with a GP, because the safety logic has to hold even in a demo. This is exactly why a human owns the call on what ships.
The guardrails I held the design to
1. Ponte never diagnoses or prescribes. It guides, translates, and hands off to a professional.
2. Structured input is the spine, not free text, so translation stays safe and consistent.
3. Red-flag symptoms route to 112, never to a booking.
4. Health information is shared only behind an explicit consent gate.
5. Data limits, like an unknown English level or an unlisted price, are shown honestly, not hidden.
06
Refining
Refining is the last mile, and it is where being a designer actually shows. There are two jobs. First, I reconcile every mismatch and error the earlier passes left behind: where the Stitch prototype and the Lovable build diverged, the invented Milan clinics that had to become Rome, the inconsistent icons, the safety slip from the critique. The fast tools hand you most of a screen. Closing the last stretch is the work.
​
Second, I rebuild the design system to be scalable. Stitch exports its system as a flat image, not as usable styles, so I rebuilt it in Figma as real tokens and components: color, the type scale, the routing system. Built this way, every new screen inherits the same rules without re-deciding them, and an engineer can build straight from the file. AI handed me the picture. Turning it into a system that survives scale and handoff is the work only a designer does.
WHAT I LEARNED
AI as intern, me as lead.
The tools gave me speed and a strong first draft. Everything that made the product trustworthy, the safety call, the honest data, the rebuilt system, came from judgment, not generation. In a world where everyone can prompt, that judgment is the job.