• 

  Landing page

• 

  Voice call start page

• 

  Voice call running

• 

  Chat landing page

• 

  Chat results

• 

  Red alerts for patients

• 

  Red alerts for clinics in Provider Portal

🤰 MaaSwara

A Mother’s Voice — Multimodal AI Triage Engine for Low-Resource Maternal Healthcare

---

🚀 From Idea to Impact

The Problem — The Delay That Kills

Every two minutes, a woman dies during pregnancy or childbirth.

• 80%+ deaths are preventable
• 94% occur in low-resource regions
• 50%+ caused by “Type 1 Delay” — failure to recognize danger signs
  

A mother experiences:

• Severe headache
  
• Blurred vision
  
• Swelling
  

She assumes it’s normal.

It is preeclampsia — and without intervention, it can kill her within hours.

Healthcare systems assume:

• Literacy
  
• Awareness
  
• Internet access
  

She has none.

---

The Solution — MaaSwara

Voice → AI → Clinical Decision → Real-Time Intervention

MaaSwara is a voice-first, multilingual AI triage system that:

• Listens in any language
  
• Understands symptoms clinically
  
• Classifies severity
  
• Acts instantly
  

No typing. No literacy. No delay.

---

🧠 What It Does

Intent → Action Mapping

Intent	Example Input	System Output
Normal symptom	“Thoda dard hai”	GREEN → reassurance
Critical symptom	“Severe headache + blurred vision”	RED → clinic alert
Fetal concern	“Baby not moving”	YELLOW → urgent follow-up
Emergency	“Heavy bleeding”	RED → override → dispatch

---

🌐 Accessibility Layers

Mode	Environment	Description
🎙️ Voice	Low literacy	Real-time speech interaction
💬 Web Chat	Standard users	UI-based interaction
📲 Telegram	2G networks	Ultra low-bandwidth input

All inputs → Single Unified Triage Engine

---

🏗️ Architecture

System Overview

┌────────────────────────────────────────────┐
│            Client Interfaces               │
│  🎙️ Voice UI  💬 Web Chat  📲 Telegram Bot │
│  (Web Audio)   (React UI)   (2G Input)     │
└─────────────────┬──────────────────────────┘
                  │ HTTP / WebSocket / Webhook
┌─────────────────▼──────────────────────────┐
│         Next.js Edge Backend               │
│  /api/chat   /api/live-token               │
│  /api/alerts /api/telegram/webhook         │
└───────────────┬───────────────┬────────────┘
                │               │
        ┌───────▼───────┐ ┌─────▼────────────┐
        │ Gemini-2.5    │ │ Deterministic    │
        │ Flash AI      │ │ Safety Engine    │
        │ (Primary)     │ │ (WHO Rules)      │
        └───────┬───────┘ └─────┬────────────┘
                │               │
                └───────┬───────┘
                        ▼
            ┌────────────────────────┐
            │   Severity Resolver    │
            │  (Override Logic)      │
            └──────────┬─────────────┘
                       │
            ┌──────────▼─────────────┐
            │   Supabase Database    │
            │   (alerts + routing)   │
            └──────────┬─────────────┘
                       │
            ┌──────────▼─────────────┐
            │ Geolocation Routing    │
            │ (Haversine Distance)   │
            └──────────┬─────────────┘
                       │
            ┌──────────▼─────────────┐
            │   Realtime Alerts      │
            │   (WebSockets)         │
            └──────────┬─────────────┘
                       │
            ┌──────────▼─────────────┐
            │   Provider Dashboard   │
            │   (Clinic Interface)   │
            └────────────────────────┘

---

⚙️ Processing Pipeline

User Input
    │
    ├──► AI Engine (Gemini-2.5-Flash)
    │         │
    │         └──► summary_en (proxy output)
    │
    ├──► Deterministic Safety Engine
    │         └──► regex match (WHO danger signs)
    │
    └──► Severity Resolver
              │
              └──► RED → Alert Dispatch

---

🔁 Agent-Level Flow

User Input
    │
    ├──► triage_engine        → classifies severity (LLM)
    ├──► safety_engine        → validates danger signs (regex)
    └──► resolver_engine      → computes final severity

If RED:
    └──► routing_engine       → finds nearest clinic
         └──► alert_engine    → pushes realtime alert

---

🧠 Core Breakthrough — AI Safety System

Problem: LLM Hallucination = Risk

AI can:

• Miss critical symptoms
  
• Misclassify severity
  

In healthcare → unacceptable.

---

Solution: Dual-Layer Decision Engine

Layer	Role
AI (LLM)	Understanding + translation
Deterministic Engine	Safety enforcement

---

📐 Mathematical Model

Let:

• ( S ) = raw input
  
• ( E ) = English proxy summary
  
• ( \mathcal{T} = {T_1, T_2, ..., T_{11}} ) = danger signs
  

Trigger condition:

$$ \exists T_i \in \mathcal{T} \mid \text{match}(T_i, S) \lor \text{match}(T_i, E) $$

Final decision:

$$ \text{Final Severity} = \max(\text{LLM Output}, \text{Deterministic Override}) $$

---

🌍 Proxy Language System

Instead of writing rules for every language:

• AI generates summary_en
• System scans only English
  

Result:

• 100+ language support
• No translation overhead
• Full safety guarantee

---

🔄 System Workflows

🎙️ Voice Workflow

Audio Capture (PCM16)
    │
    ├──► WebSocket Stream
    │
    ├──► Gemini Live Processing
    │
    ├──► Safety Override Check
    │
    └──► Severity Output → Alert if RED

---

📲 Telegram Workflow

User Message
    │
    ├──► Webhook (Next.js API)
    │
    ├──► AI Processing
    │
    ├──► Markdown Stripped
    │
    └──► Response Delivered

---

🚨 Alert Dispatch Workflow

RED Detected
    │
    ├──► Capture GPS
    ├──► Haversine Distance Calculation
    ├──► Assign nearest clinic
    ├──► Store in Supabase
    └──► Push via WebSocket → Dashboard

---

🛠️ Tech Stack

Layer	Technology
Frontend	React + Next.js
AI Engine	Gemini-2.5-Flash
Voice	Web-Audio-API + AudioWorklet
Backend	Next.js Edge APIs
Database	Supabase PostgreSQL
Realtime	WebSockets
Routing	Haversine Algorithm
Hosting	Vercel

---

⚔️ Challenges

• LLM hallucination risk → solved via deterministic override
  
• Multilingual scaling → solved via proxy English system
  
• Web Audio constraints → manual activation fix
  
• Telegram API failures → raw text fallback
  
• Real-time reliability → WebSocket pipeline
  

---

🏆 Accomplishments

• Built a real-time AI healthcare system
• Achieved 100+ language support
• Designed zero-retention voice pipeline
• Created end-to-end emergency workflow
• Solved AI safety in critical environments
• Delivered a fully working system

---

📚 What We Learned

• AI must be controlled, not trusted blindly
• Real-world impact requires system-level thinking
• Accessibility > complexity
  
• Constraints drive innovation
  
• Voice is the most powerful interface in low-resource systems
  

---

🚀 What's Next

• EHR / DHIS2 integration
  
• Biometric device integration
  
• WhatsApp & IVR support
  
• Automated ambulance dispatch
  
• Government-level deployment
  

---

🌍 Impact

MaaSwara transforms maternal healthcare from:

❌ Reactive
❌ Delayed
❌ Literacy-dependent

To:

✅ Real-time
✅ Voice-first
✅ Accessible
✅ Life-saving

---

✨ Final Statement

"Listen to the mother. Override the hallucination. Save the life."

Built With

• gemini-2.5-flash
• next.js
• postgresql
• react
• supabase
• typescript
• vercel
• web-audio-api
• websockets