CarePath AI

• 

  ClearPath Ai Logo

Inspiration

Our team is based in Canada, and we’ve seen the same pattern show up here as in the U.S.: when people feel unwell, it’s surprisingly hard to know what level of care they actually need. Many patients wait too long because they don’t want to “waste anyone’s time,” while others go straight to the emergency department because it feels like the safest option. The problem is even worse in rural and underserved communities, where access is limited and a single wrong decision can mean hours of travel, long waits, and delayed treatment.

This uncertainty creates a system-level issue: non-urgent cases can flood higher-acuity settings, urgent patients face longer delays, and clinics and hospitals spend time re-triaging or redirecting people who simply went to the wrong place. We were inspired to build a better front door to healthcare: a simple, human-feeling system that helps patients quickly and confidently choose the right next step while reducing avoidable strain on already stretched facilities.

We also grounded our idea in real clinical perspective. We spoke with two registered nurses and one medical school student, and shaped our flow and outputs around their feedback on what clinicians actually need to see quickly.

What it does

CarePath AI is an online voice-based intake assistant. Using real-time voice (via LiveKit), it talks with the patient, captures key details like symptoms, timeline, medications, allergies, and context, and turns that conversation into a structured intake summary a clinician can review fast. The clinician then decides the best next step, such as: Treat at home with guidance, Go to a local clinic, Go to a hospital, Schedule a call or virtual consultation.

The patient then receives a notification outlining the recommended course of action, streamlining them into the right level of care without having to repeat their story multiple times.

How we built it

We designed the user experience first in Figma to keep the interface clean, trustworthy, and easy to use for both patients and clinicians. We mapped the full workflow in Miro to define how information moves from patient voice input to a clinician-ready summary. We built the front end in Cursor, then integrated the back end by connecting LiveKit for real-time audio sessions and transcription and incorporating our integration stack (including LeanMCP) to support the pipeline from conversation to structured output.

Based on feedback from the nurses and medical student, we focused on keeping patient questions simple and non-technical, producing a clinician-facing summary that’s fast to scan, highlighting key safety details (timeline, potential red flags, meds/allergies), and ensuring the workflow supports a clear “next step” decision.

Challenges we ran into

The hardest part was integration: connecting the voice stack, transcription, UI, and structured output into one smooth experience. Managing API keys was a constant blocker, especially when tools required paid tiers or had restrictions that didn’t fit student plans, so we had to work around limitations while keeping things secure. We also ran into inconsistent documentation and setup steps across providers, which forced us to learn new tools quickly. Debugging was another major challenge because one small configuration issue could break the entire pipeline. On top of that, delays with sponsor track approval meant we had to stay flexible and keep iterating without losing momentum.

Accomplishments that we're proud of

We’re proud that we took CarePath AI from idea to a working end-to-end prototype with a real patient flow and a clinician view. We built a polished interface that feels like a real healthcare product, not a hacky demo, and we created a workflow that turns a messy conversation into something a doctor can actually use.

We’re also proud of how much we iterated. Even with constraints and delays, we narrowed broad ideas into a clear need, broke features down until the workflow was solid, and improved the product based on real stakeholder feedback rather than assumptions.

What we learned

This was our first time building an AI agent end-to-end, and we learned quickly that a workflow that looks perfect on paper can break in real implementation. We learned how to build a voice-first product with LiveKit, how much a clean design system in Figma matters when iterating fast, and how Cursor speeds up development loops and debugging. Using Miro helped us think more clearly about architecture and system flow before writing code, which saved time later.

What's next for CarePath AI

Next, we want to pilot CarePath AI with real clinics and build a direct connection to doctors so they can recommend it to their patients as the first step before visits. We also want to improve clinician tools like review and editing workflows, follow-up prompts, and clarification requests.

On the patient side, we want to add a map that shows the closest appropriate place to go (walk-in clinic, urgent care, or hospital) based on the clinician’s decision. Longer term, we want CarePath AI to fit directly into clinic workflows so intake becomes a standard, reliable process that saves time for providers and gets patients to the right care faster.

Built With

• cursor
• figma
• leanmcp
• livekit
• python