MediRoute

Inspiration

What it does

How we built it

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for MediRoute

Inspiration

We kept coming back to a simple, frustrating moment: someone feels unwell, and the first real obstacle isn't the illness — it's figuring out where to go. Is this a GP visit or a specialist? Is the nearest clinic even open right now? Does it take my insurance? Search engines and hospital directories hand back a list and leave the deciding to you. We wanted to build something that actually helps make that decision — without crossing the line into telling someone what's wrong with them. That last constraint, straight from the ZERO ORIGIN problem statement — navigation and coordination, never diagnosis — became the design principle we built everything else around.

What it does

MediRoute is an agentic AI that listens to a patient's need (voice or text) and walks them from confusion to a booked appointment:

Urgency gate — the first question is always "is this an emergency?" If yes, it skips everything else and routes straight to the nearest ER. Non-diagnostic triage — follow-up questions map the complaint to a service category (e.g. "skin issue" → dermatology), never to a disease guess. Ranked facility matching — scores nearby facilities on distance, open-now status, accessibility, insurance, and language, and shows why each one was suggested. Live navigation — real-time directions via the Google Maps API. Agentic booking — proposes open slots, confirms the time with the patient, and books it on their behalf. How we built it

We designed the system around five layers: an input/understanding layer (speech-to-text + NLP intent extraction), an LLM-driven reasoning engine running the urgency → category → filter question tree, a facility-matching layer that scores and ranks candidates, a navigation layer on the Maps API, and a booking layer we scoped as a mocked scheduling backend, since live hospital API access wasn't realistic at hackathon stage. All of it sits behind a React front end, with a 2D lip-sync voice avatar as the patient-facing interface.

Challenges we ran into Staying on the right side of the diagnosis line. It's a fine distinction — "do you have a fever?" is diagnostic, "is this urgent or can it wait?" is navigational — and we had to consciously redesign our conversation flow more than once to stay compliant with the problem statement's constraint. No live hospital data or booking APIs. We had to design a convincing mocked scheduling layer while being explicit that it's a simulation, not a live integration — a trade-off we think is honest rather than a shortcut. Balancing the avatar with the substance. It was tempting to let the voice/avatar interface be the star of the pitch, but we had to keep reminding ourselves that the real value is in the reasoning and matching logic underneath it. Accomplishments that we're proud of Designing a non-diagnostic question tree that still feels genuinely helpful, not like a form to fill out. The emergency-override design — urgency detection doesn't just reprioritize the ER in a list, it disables the booking flow entirely and redirects immediately, which we think is the detail that shows we took the real-world stakes seriously. Making the facility ranking explainable — every suggestion comes with the reasons behind it (distance, hours, accessibility, insurance) instead of a black-box "here's your answer." Scoping an agentic loop — listen, reason, decide, act — that stays coherent end-to-end, from a first spoken sentence to a confirmed appointment. What we learned The hardest part wasn't the AI — it was the guardrails. Getting an agent to ask genuinely useful follow-up questions without drifting into symptom-based diagnostic reasoning took several redesigns of our question tree. Explainability builds trust. Early versions of our facility ranking were a black box. Surfacing why a facility was suggested turned out to matter as much as the ranking itself. Safety-by-default is a feature, not an afterthought. Small design choices — like disabling booking outright on an emergency signal — carry outsized weight in how trustworthy the whole system reads. What's next for MediRoute Integrating with real hospital scheduling APIs in place of the mocked booking backend. Multilingual voice support for wider regional accessibility. Upgrading the interface to a 3D avatar for a more immersive coordination experience. Insurance-provider integration for real-time coverage checks during facility matching. A feedback loop — post-visit check-ins — to improve facility ranking accuracy over time.

Built With

• agentic-ai
• api
• conversational-ai
• express.js
• firebase
• google-maps
• groq
• javascript
• llm
• natural-language-processing
• node.js
• openai-api
• react
• speech
• speech-recognition
• speech-to-text
• text
• to