CarePath

Inspiration

When you feel sick, the hardest question is often the first one: where do I even go? Emergency room, urgent care, your doctor, telehealth, or just rest at home? Pick wrong and you either waste hundreds of dollars or risk your health. Most people guess. Made to give patients one place that is actually theirs, that listens, explains its thinking, and puts them in control of their own care instead of leaving them lost in a phone tree.

What it does

CarePath is a voice first personal healthcare center. You talk about what is going on and it points you to the right level of care, estimates what it may cost with your insurance, and builds a shareable Care Card with the reasoning, red flags, what to say at check in, and what to bring.

It has five connected modes:

• Triage: describe symptoms, get a care recommendation with a confidence score and cost estimate.
• Debrief: just left the doctor, get a plain language explanation of what they said and your next steps.
• MedCard: speak your medications and allergies, get an interaction check, or scan a pill bottle with your camera.
• Check in: a short voice mental health check in that notes what to raise with your provider.
• Timeline: log symptoms over time, which feeds straight into your triage.

It can pull your existing health data through EPIC MyChart (325+ million real patients' data profiles), and it can generate a Care Card from whatever data you already have, even just your symptom log. The Care Card can also surface relevant patient communities so you can hear from people with similar experiences, always with clear disclaimers. Everything stays in your browser.

How we built it

CarePath is a Next.js 16 app with React 19, TypeScript, and Tailwind v4, deployed on Vercel. Live voice runs on Grok Voice over the xAI Realtime API, which can fact-check on the web while you speak. Use OpenAI gpt-4o-mini with structured JSON output for the triage classifier, the conversation modes, the community matcher, and the camera pill bottle reader using its vision capability. Relevant Reddit communities are pulled and verified through the public arctic_shift API. Insurance and pricing are synthetic, and all patient data lives in localStorage with no database, so nothing leaves the device unless the user chooses.

Challenges we ran into

Getting real time speech to speech working took the most effort, from the correct Grok Voice endpoint and WebSocket handshake to keeping the transcript clean and letting the agent speak first. On the safety side, the emergency detector first false triggered on negations like "no chest pain," and an early version could quietly downgrade a real emergency when an API call failed, so I made it fail safe toward the emergency room. I also fought a few classic front end bugs, including an infinite render loop from an external store and a content security policy that blocked the dev build.

Accomplishments that we're proud of

A complete voice loop that shows its reasoning and a confidence level instead of a black box answer. Care recommendations that are cost aware and tied to your insurance plan. A Care Card you can generate from any of your data. A working camera pill bottle scanner. Patient community discovery with honest disclaimers. And a product that keeps your data private, never places a real 911 call, and always tells you it is a navigation tool, not a diagnosis system.

What we learned

Safe medical AI is mostly about what happens when things go wrong, so I learned to design every fallback to protect the patient. Plain language beats clinical jargon for anxious users.

What's next for CarePath

Multi language support, a validated drug interaction database, scheduling and pharmacy handoff, and caregiver sharing.

Built With

• arctic-shift-api
• gpt-4o-mini
• gpt-4o-mini-vision
• grok-voice
• next.js
• node.js
• openai
• react
• smart-on-fhir
• tailwindcss
• typescript
• vercel
• web-audio-api
• xai-realtime-api