Cura

Inspiration

Many older people take several medicines a day, and the instructions are often handwritten in a way almost nobody can read. Families worry whether a dose was taken, and the person taking the medicine is rarely told in plain words what each tablet is for. We wanted to make that daily routine calmer and more independent, for seniors and for the people who care about them.

What it does

Cura turns a photographed prescription into a clear daily plan.

• Scan: photograph or upload a prescription, and Gemini reads it.
• Review: a caregiver checks every medicine side by side with the original photo. Anything the AI is unsure about is flagged "Check this", and nothing is saved until it is confirmed.
• Four simple parts of the day: Dawn, Midday, Dusk and Night, with large text and big tap targets.
• Understand: each medicine is explained in plain language ("commonly used to...", "how it helps"), in English, Hindi or Marathi, and can be read aloud.
• Tell family: after a dose, one tap opens WhatsApp with a prefilled message to a saved family contact.

Cura organises and explains. It never diagnoses, recommends or changes doses.

How we built it

• Design: the interface was designed with Google Stitch, then cleaned up so it follows one colour, type and spacing system, and works on mobile and desktop.
• App: a responsive web app built with Vite, Tailwind CSS and vanilla JavaScript, developed with Google Antigravity.
• AI: two Gemini calls. The first reads the prescription image into structured JSON. The second explains each medicine using only its generic name, never the image or any patient detail.
• Safety by design: shorthand such as "1-0-1" or "BD" is decoded in our own code, not by the AI. Names are matched against a local medicine list. A human confirms every schedule, and explanation text is validated so it can't contain dosing advice.
• Backend: small serverless functions keep the API key off the browser. Schedules and dose history stay in the user's own browser storage.

Challenges we ran into

• Handwriting is hard. Medicine names are easier to read than instructions, so we built confidence flags, an editable review screen and a manual-entry fallback instead of trusting the AI blindly.
• Medical content needs guardrails. We had to remove hardcoded claims from our first design and constrain what Gemini may say.
• Free-tier limits. We designed around small request quotas with caching, a fallback model and a rehearsal mode.
• Cluttered first designs. Our first generated UI tried to show everything at once, so we rewrote the prompts to one screen at a time with strict rules.
• WhatsApp links need a final tap. We say so in the app instead of overpromising "automatic alerts".

What we learned

• For health tools, "AI reads it, a person confirms it" is a better design than chasing accuracy.
• Doing decoding and validation in plain code makes the AI's mistakes easier to catch.
• Simple, calm design matters more than features for older users.

What's next

• Scan a medicine's label to check it matches the schedule.
• Adherence history and weekly stats.
• Automatic family notifications and shared caregiver access.
• Verified medicine information sources and more Indian languages.

Built With

• css3
• fuse.js
• gemini-api
• getusermedia
• google-ai-studio
• google-antigravity
• google-stitch
• html5
• javascript
• localstorage
• material-symbols
• node.js
• nunito-sans
• tailwind-css
• vercel
• vercel-serverless-functions
• vite
• web-audio-api
• web-speech-api
• whatsapp