FEVERAI

AI-powered tropical fever triage for rural clinics.

Comment

Inspiration

Every monsoon, rural India faces deadly waves of dengue, malaria, typhoid, and chikungunya. A simple fever can turn fatal because frontline health workers (ASHAs, ANMs, and PHC doctors) have almost no tools to tell these diseases apart quickly. Labs are hours away, RDT kits are in short supply, and by the time a correct diagnosis is made, the patient is often critical—and outbreaks have already started spreading silently.
We wanted to put a tropical-medicine specialist in every rural health worker’s pocket using nothing more than a smartphone and AI.

What it does

FeverAI is an end-to-end AI-powered tropical fever triage + surveillance platform that turns any Android/iOS phone into a smart diagnostic hub. Health workers simply:

  • Fill a quick symptom form
  • Snap photos of rashes, tongue, eyes, or Rapid Diagnostic Tests
  • Record a cough (optional)
  • Or start a live AI-guided video exam

Our Gemini 1.5 Pro multimodal engine instantly returns a structured report with primary diagnosis, differentials, risk score, and clear next steps—while anonymously pushing the case to a real-time district surveillance dashboard that also correlates fever spikes with climate data for early outbreak alerts.

How we built it

  • Frontend: React + Vite + Tailwind CSS for a fast, offline-capable PWA
  • Backend & AI: Google Gemini 1.5 Pro (multimodal reasoning) + 1.5 Flash (TTS for voice guidance) via Vertex AI
  • Live video diagnostics: Custom WebRTC + Gemini tool-calling to capture images mid-call
  • Surveillance dashboard: Firebase + Google Cloud Run + BigQuery for real-time aggregation and visualization
  • Deployment: Fully hosted on Google Cloud Run (zero-server management in 48 hours!)

All built from scratch during the Micro Labs Hackathon 2025.

Challenges we ran into

  • Getting reliable multimodal outputs (image + text + audio) from Gemini in a single prompt — solved with careful system-prompt engineering and JSON schema enforcement
  • Live video frame extraction on low-end devices — worked around using getUserMedia + canvas snapshots
  • Keeping the app under 5-second response even on 2G — aggressively cached models and used Gemini-Flash for lighter tasks
  • Ensuring medical accuracy and transparency — built an “AI Training Hub” so users can see and improve the knowledge base

Accomplishments that we're proud of

  • Fully functional multimodal diagnosis (symptoms + rash + RDT + cough audio) working in <6 seconds on a ₹8,000 phone
  • Live AI doctor that verbally guides physical examination in real time (Hindi + English TTS)
  • Real-time district-level outbreak dashboard with climate correlation — something most state health departments still don’t have
  • Transparent & improvable AI: health workers can train the model with new labeled images directly from the app
  • 100% open-source and built in just 48 hours!

What we learned

  • Modern multimodal LLMs are now good enough to outperform traditional rule-based clinical decision tools for tropical fevers
  • Structured JSON output + tool-calling makes AI behave like a reliable clinical API
  • Rural health workers love voice + visual guidance more than forms — live video consultation was the #1 requested feature during our user testing
  • Privacy-by-design matters: all patient data stays on-device unless explicitly shared for surveillance

What's next for FeverAI

  • Pilot with 100 ASHAs in Odisha & Jharkhand starting January 2026
  • Add offline-first mode using Gemini Nano (on-device)
  • Integrate with ABDM (Ayushman Bharat Digital Mission) for seamless patient records
  • Expand to Africa & Southeast Asia (already training on local datasets)
  • Partner with state health departments to make the surveillance dashboard official

FeverAI is just the beginning—AI can finally bring specialist care to the last mile.

MicroLabsHackathon #BuiltWithGemini #HealthAI

Built With

Share this project:

Updates