praveenkumar-dev-art/gemma-4-competition

Purpose and entry points

Sanjeevani is a multilingual, voice-first rural-health triage application. The primary web entry is sanjeevani-app/src/main.jsx, which mounts sanjeevani-app/src/App.jsx. It is built as a React/Vite PWA and can be packaged into Android through Capacitor; the Android host starts at sanjeevani-app/android/app/src/main/java/com/sanjeevani/triage/MainActivity.java.

Application shell and safety UI

App.jsx composes the user journey and persistent safety surfaces: disclaimer, header, intake, chat/triage, emergency handling, and reports.

• sanjeevani-app/src/App.jsx

• sanjeevani-app/src/components/Disclaimer.jsx

• sanjeevani-app/src/components/EmergencyModal.jsx

The emergency UI is a control boundary: RED triage escalates to a blocking 108-call overlay rather than normal conversational continuation.

Patient intake and interaction layer

The app collects patient demographics, selected language, and symptom conversation before or during triage. Voice-oriented interaction is implemented in dedicated UI components, while chat renders bilingual clinical guidance.

• sanjeevani-app/src/components/PatientIntakeWizard.jsx

• sanjeevani-app/src/components/VoiceWidget.jsx

• sanjeevani-app/src/components/ChatContainer.jsx

Input crosses from browser/device interaction into shared clinical state; the design supports Tamil, Malayalam, Telugu, Hindi, and English.

Triage state and agentic memory

TriageContext is the central client-side state boundary. It accumulates patient and conversation information across turns, including the documented Patient, Rough, Clean, and concentrated-question memory pages. UI components consume this shared state rather than independently determining care urgency.

• sanjeevani-app/src/context/TriageContext.jsx

• sanjeevani-app/src/components/TriagePanel.jsx

Clinical reasoning pipeline and offline fallback

Clinical logic applies SOCRATES-guided questioning, a symptom ontology, and WHO-aligned severity classification. This local engine is the continuity path when no model API key or network is available; it should remain independent from remote inference.

• sanjeevani-app/src/scripts/clinicalAgent/index.js

• sanjeevani-app/src/scripts/clinicalAgent/clinicalReasoningEngine.js

• sanjeevani-app/src/scripts/clinicalAgent/symptomOntology.js

The outcome controls three principal routes: RED emergency escalation, YELLOW clinic referral, and GREEN self-care guidance.

Remote Gemma inference boundary

Prompt construction and model access are separated from the UI and deterministic clinical fallback. Prompt templates constrain multilingual SOCRATES dialogue and structured output; the service calls Google’s Gemini API for Gemma-compatible model responses, configured through Vite environment variables.

• sanjeevani-app/src/scripts/clinicalAgent/doctorPrompts.js

• sanjeevani-app/src/services/gemmaService.js

Remote output must be interpreted into application triage state, not treated as unbounded presentation content.

Care routing, reports, and external runtime services

Triage results drive clinic discovery/referral, patient-facing triage panels, and a dual report/pass surface. Mock/local data supports development or offline scenarios.

• sanjeevani-app/src/components/ClinicLocator.jsx

• sanjeevani-app/src/components/TriagePass.jsx

• sanjeevani-app/src/data/mockData.js

A separate Node HTTP runtime provides /api/tts audio support on port 3001: sanjeevani-app/server.js. Web builds use Vite (sanjeevani-app/vite.config.js); Android packaging uses Capacitor (sanjeevani-app/capacitor.config.json and sanjeevani-app/android/). CI runs through .github/workflows/ci.yml.