HealthNav

Inspiration

Healthcare guidance is often hardest to access when people are stressed, unsure how serious their symptoms are, or more comfortable speaking in their native language. We built HealthNav to make first-step healthcare navigation faster and more accessible for people in India by combining multilingual symptom intake, conservative AI triage, and nearby facility discovery in one lightweight experience.

What it does

HealthNav is a multilingual AI healthcare triage and facility finder agent. A user can describe symptoms by text or voice, choose Auto / English / Hindi / Marathi, and receive a structured triage result with:

• urgency level: Emergency / Urgent / Routine
• probable conditions
• recommended next action
• red flags to watch for

After triage, HealthNav uses geolocation plus OpenStreetMap data to surface nearby hospitals and clinics on an interactive map. If the case looks critical, the app prominently escalates with one-tap emergency actions for 108 and a WhatsApp share flow. How we built it

1. Intake Agent collects symptoms from text or browser voice input and detects/prepares the working language.
   1. Triage Agent sends structured symptom context to NVIDIA NIM using meta/llama-3.1-70b-instruct and requests strict JSON output.
   2. Facility Finder Agent uses browser geolocation and OpenStreetMap / Overpass data to find nearby care options.
   3. History Agent keeps an in-memory session timeline and allows export as a JSON report.
   4. Escalation Agent highlights emergencies and prepares rapid-contact actions.

The frontend is a single no-build HTML/CSS/JS app, which keeps the project easy to audit, easy to deploy from GitHub, and fast to demo. We deployed the live demo on Vercel and kept the repo on GitHub.

Challenges we ran into

One of the biggest challenges was making a static-first app work reliably with live AI inference. Direct browser calls to NVIDIA NIM ran into CORS restrictions, so we built a lightweight serverless proxy on Vercel to securely forward triage requests without exposing secrets in the repo.

Another challenge was reliability for nearby facility search. Public Overpass endpoints can occasionally timeout, so we added a same-origin facility proxy with retries across Overpass mirrors to keep the live demo stable during judging.

We also spent time hardening the demo UX:

• graceful fallback handling for every fetch
• strict JSON parsing and malformed-output recovery for LLM responses
• clear messaging on GitHub Pages so reviewers are guided to the live AI deployment
• non-intrusive emergency escalation that surfaces actions without force- opening phone or WhatsApp apps ## What we learned We learned that “agentic” is not just about chaining API calls. A useful real-world agent needs orchestration, defensive parsing, fallback logic, deployment-aware UX, and clear escalation behavior. We also learned how much trust and usability matter in healthcare-adjacent experiences: conservative outputs, explicit red flags, and understandable next steps are just as important as model quality. ## What's next for HealthNav Next, we’d like to add:
• richer multilingual support beyond English, Hindi, and Marathi
• accessibility improvements for low-literacy users
• structured medication / allergy intake
• saved patient profiles with secure backend storage
• hospital specialty filtering and better care routing
• clinician-facing handoff summaries

Built With

• 3.1
• 70b
• api
• css
• github
• gps
• html
• instruct
• javascript
• leaflet.js
• llama
• nim
• nvidia
• openstreetmap
• overpass
• pages
• speech
• vercel
• web