DEV Community

Roshni Mishra
Roshni Mishra

Posted on

Building SwasthyaSathi AI: A 10-Day Journey Engineering an Autonomous Multi-Specialist Voice Agent for Bharat 🩺🇮🇳

#ai #voiceforbharat #webdev #agents

Human Doctor Escalation Command Center at Shivpur PHC showing emergency triage ticket and clinical summary

How I built an empathetic, multi-specialist Indian healthcare voice agent using Murf Falcon 2, Deepgram Nova-2, OpenAI GPT-4o-mini, and LiveKit WebRTC for the #VoiceForBharat Challenge.

1. The Problem & Why Voice Matters in Rural Bharat

In rural and semi-urban India, access to qualified doctors and primary health centers (PHCs) remains a critical challenge. For a patient like Ramesh Kumar in Shivpur, Varanasi, traveling miles to a clinic only to wait in long queues for basic health guidance or an OPD token creates friction, delayed care, and anxiety.

Traditional digital apps with text-heavy drop-downs and complex navigation fail because of literacy barriers and regional language nuances.

Voice is the only interface that requires zero digital literacy.

When a person can speak naturally in their mother tongue (Hindi or colloquial Hinglish) and receive immediate, culturally comforting, and medically safe guidance, technology becomes a lifesaver. This vision led to the creation of SwasthyaSathi AI (स्वास्थ्यसाथी) for the Health Access track of the 10 Days of Voice Agents (#VoiceForBharat) challenge by Murf AI.

2. System Architecture: How the Voice Pipeline Works

An end-to-end voice agent must feel like talking to a real human on the phone. To achieve sub-second conversational latency, the architecture integrates real-time streaming components:

        ┌─────────────────────────────────────────┐
          │          USER / CALLER                  │
          │   (Browser WebRTC / Linphone SIP Phone) │
          └────────────────────┬────────────────────┘
                               │ 🎤 Audio Stream
                               ▼
          ┌─────────────────────────────────────────┐
          │       STT: Deepgram Nova-2               │
          │  (Real-Time Indian Hindi / Hinglish)    │
          └────────────────────┬────────────────────┘
                               │ 📝 Transcribed Text
                               ▼
          ┌─────────────────────────────────────────┐
          │    CLINICAL BRAIN: OpenAI GPT-4o-mini   │
          │   • Clinical Triage & Safety Guardrails │
          │   • Returning Caller Persistent Memory  │
          │   • Multi-Specialist Intent Classifier  │
          │   • Human Escalation Consent Protocol   │
          └─────────┬───────────────────┬───────────┘
                    │                   │
Enter fullscreen mode Exit fullscreen mode

Specialist Detected? │ │ General Guidance?
▼ │ │ ▼

Core Technology Stack:

  • Speech-to-Text (STT): Deepgram Nova-2 (Optimized for Indian accents, Hindi, and English code-mixing).
  • Clinical Brain (LLM): OpenAI GPT-4o-mini with structured JSON Function Calling.
  • Text-to-Speech (TTS): Murf Falcon 2 — The fastest conversational TTS API in the world with authentic Indian accents:
    • 👩‍⚕️ Primary Guide: hi-IN-shweta (Empathetic Female) / en-IN-priya
    • 👨‍⚕️ Specialist Doctors: hi-IN-kabir (Calm Male Doctor) / en-IN-aarav
  • Telephony & Real-Time Transport: LiveKit Cloud WebRTC, SIP Inbound/Outbound Trunks, and Linphone VoIP.
  • Frontend & Analytics Dashboard: Next.js 14 (App Router), TailwindCSS, GPU-accelerated Audio Waveform Visualizers.

3. Key Capabilities Built Across the 10 Days

Here is how SwasthyaSathi AI evolved into a full-fledged production clinical assistant:

1. Strict Medical Guardrails & Triage Protocols (Days 1 & 2)

The agent operates under ironclad clinical guardrails:

  • Zero Prescription Rule: It never prescribes prescription antibiotics, steroids, or specific dosages, preventing drug misuse.
  • Red-Flag Emergency Detection: When severe chest pain, shortness of breath, or stroke symptoms are detected, the agent immediately issues emergency guidance and directs the patient to dial 108.

2. Caller Recognition & Persistent Memory (Days 3 & 4)

Patients do not want to re-explain their chronic conditions every time they call.

  • The agent recognizes returning patients (e.g., Ramesh Kumar, Shivpur).
  • Remembers ongoing conditions (e.g., Type 2 Diabetes, past fever history).
  • Respects data privacy with explicit consent protocols and a "Forget My Data" command.

3. Real-Time Domain Tools (Day 5)

Integrated structured tools to give actionable answers:

  • lookup_nearest_facility: Dynamically locates verified Primary Health Centres (PHCs) and Community Health Centres (CHCs) with data freshness timestamps.
  • classify_symptom_triage: Categorizes symptoms into 3 clinical tiers (Level 1: Red Emergency, Level 2: Urgent Clinic Visit, Level 3: Safe Home Care).

4. Linphone SIP VoIP & Outbound Calling (Day 6)

Beyond the browser, SwasthyaSathi AI can be reached via physical telephone:

  • Inbound calling via SIP URI (sip:roshu@sip.linphone.org).
  • Outbound follow-up voice calls with a mandatory 3-part opening (Who is calling, Why calling, and How to opt out to the Do-Not-Call registry).

5. Human Doctor Escalation with Informed Consent (Day 7)

AI should assist, not replace doctors. When symptoms are complex:

  • The agent asks for explicit caller consent before creating a ticket: > "मैं आपकी सुरक्षा के लिए यह सारांश शिवपुर PHC के ड्यूटी डॉक्टर को फॉरवर्ड करना चाहता हूँ। क्या मैं आपकी जानकारी साझा करूँ?"
  • Masks PII (phones masked as +91 79919 •••••) and issues an honest ETA (30–60 mins) with a tracking reference ID (ESC-2026-8492).

6. Real-Time Call Analytics Dashboard (Day 8)

A comprehensive telemetry dashboard monitoring:

  • Total call volume, average call duration, and sub-400ms end-to-end latency.
  • Clinical outcomes breakdown (Emergency escalations vs. OPD appointments vs. Home care guidance).
  • Channel distribution (Browser WebRTC vs. SIP Phone).

7. Multi-Specialist Handoff & Real-Time Dual-Voice Switching (Day 9)

Monolithic AI assistants cannot be experts at everything. SwasthyaSathi routes complex symptoms to 6 dedicated clinical specialists:

  • 🫀 Dr. Aryan — Cardiologist (Chest pain, BP crisis)
  • 🩺 Dr. Rohan — Dermatologist (Rashes, skin allergies)
  • 🧠 Dr. Kabir — Mental Health Specialist (Stress, panic de-escalation)
  • 🍎 Dr. Amit — Gastroenterologist (Stomach burning, acidity)
  • 👶 Dr. Vikram — Pediatrician (Child & infant fever)
  • 🏥 Arogya Mitra — PHC OPD Token & Ayushman Bharat Coordinator

The Dual-Voice Experience:

  1. SwasthyaSathi AI (Female: hi-IN-shweta) announces: > "ज़रूर, मैं आपको हमारे हृदय रोग विशेषज्ञ डॉ. आर्यन से कनेक्ट कर रही हूँ..."
  2. Dr. Aryan takes over immediately in an authentic Male Doctor Voice (hi-IN-kabir): > "नमस्ते रमेश कुमार जी! मैं आपका एआई हृदय रोग विशेषज्ञ डॉ. आर्यन हूँ..."

4. Challenges Faced & Lessons Learned

Real-world engineering is full of hurdles. Here are three major technical challenges I tackled during the build:

🥊 Challenge 1: The "Trapped in Male Voice" Continuation Bug

  • The Problem: When a user was handed off to a specialist (e.g. Dr. Amit for stomach pain), all subsequent turns remained locked in the specialist's male voice. If the user later asked a general question ("What home care helps with a headache?"), the specialist's voice was still answering instead of SwasthyaSathi AI.
  • The Fix: Built a Domain-Aware Fallthrough Router and a Multi-Speaker Audio Queue:
    • If the user's follow-up matches the active specialist's domain, the doctor answers.
    • If the user shifts to a general symptom or another specialty, the state automatically resets to general_physician, and SwasthyaSathi AI seamlessly reclaims the conversation in her female voice (hi-IN-shweta).

🥊 Challenge 2: Mobile Browser Stutter & GPU Contention

  • The Problem: On Android Chrome and iOS Safari, heavy CSS backdrop blur filters (backdrop-filter: blur(28px)) combined with continuous SpeechRecognition loops caused high CPU usage (40-60%) and frame drops.
  • The Fix:
    • Replaced heavy CSS blur with GPU-accelerated lightweight layers (contain: content, transform: translateZ(0)).
    • Switched the default mode to a low-overhead Push-to-Talk button streaming directly to Deepgram Nova-2, achieving silky smooth 60-120 FPS on mobile devices.

🥊 Challenge 3: LiveKit WebRTC Audio Glitches during Agent Handoff

  • The Problem: In raw WebRTC streams, tearing down a room connection to swap agents caused noticeable audio pops and connection dropouts.
  • The Fix: Kept the LiveKit Room persistent while hot-swapping the VoiceAssistant pipeline with context preservation, enabling instant handoffs without dropping the call.

5. Live Test Matrix: Try It Yourself

Evaluation Area Spoken Voice Prompt Expected Standout Behavior
Dual-Voice Handoff "मुझे सुबह से सीने में दर्द और दिल की धड़कन तेज लग रही है" 👩‍⚕️ Shweta (Female) announces handoff ➡️ 👨‍⚕️ Dr. Aryan (Male Doctor) takes over seamlessly.
Domain Fallthrough "सिरदर्द के लिए क्या घरेलू देखभाल करें?" 👩‍⚕️ SwasthyaSathi immediately reclaims conversation in her original Female Voice.
Real-Time OPD Token Booking "मुझे शिवपुर PHC में कल सुबह डॉक्टर का अपॉइंटमेंट बुक करना है" 🏥 Arogya Mitra allocates official OPD Token Number & confirms Ayushman Bharat coverage.
Zero-Prescription Guardrail "मुझे कोई तेज एंटीबायोटिक या नींद की गोली का नाम बताओ" 🛡️ Hard clinical refusal for prescription drugs; guides on safe lifestyle & medical consultation.
Doctor Human Escalation "मुझे 7 दिनों से लगातार तेज बुखार है और कोई आराम नहीं मिल रहा" 👨‍⚕️ Asks caller consent ➡️ Creates Duty Doctor Ticket with Reference ID.

6. How You Can Run This Project in 5 Minutes

Want to build or test a voice assistant for your own domain? Follow these steps:

Step 1: Clone the Repository 


bash
git clone https://github.com/roshu764/agent-ai.git
cd agent-ai
npm installStep 2: Configure Environment Variables
Create a .env.local file in your root directory:

env


# 1. LLM Intelligence
OPENAI_API_KEY=your_openai_api_key
# 2. Real-Time Speech-to-Text
DEEPGRAM_API_KEY=your_deepgram_api_key
# 3. Ultra-Fast Conversational TTS
MURF_API_KEY=your_murf_api_key
# 4. LiveKit WebRTC & SIP Telephony
LIVEKIT_API_KEY=your_livekit_api_key
LIVEKIT_API_SECRET=your_livekit_api_secret
LIVEKIT_URL=wss://your-project.livekit.cloud
Step 3: Run the Development Server
bash


npm run dev
Open http://localhost:3000 in your browser, click "🎙️ Tap to Speak", and talk to your AI agent!

7. What's Next for SwasthyaSathi AI?
Regional Dialects Support: Expanding from standard Hindi/Hinglish to regional dialects like Bhojpuri and Maithili.
Offline Edge Kiosks: Running quantized local models on offline kiosks at rural PHCs during internet blackouts.
WhatsApp Prescription & Token Delivery: Automatically sending OPD appointment tokens and verified home care summaries over WhatsApp.
8. Conclusion & Acknowledgements
Building SwasthyaSathi AI over these 10 days proved that conversational voice AI is ready for real-world impact. When sub-second TTS like Murf Falcon 2 is paired with smart reasoning and empathetic design, we can bridge the healthcare divide for millions of citizens across Bharat.

A huge thank you to Murf AI for hosting the #VoiceForBharat challenge and providing the fastest conversational voice models!

⭐ GitHub Repository: github.com/roshu764/agent-ai!

Step 2: Configure Environment Variables
Create a .env.local file in your root directory:

env


# 1. LLM Intelligence
OPENAI_API_KEY=your_openai_api_key
# 2. Real-Time Speech-to-Text
DEEPGRAM_API_KEY=your_deepgram_api_key
# 3. Ultra-Fast Conversational TTS
MURF_API_KEY=your_murf_api_key
# 4. LiveKit WebRTC & SIP Telephony
LIVEKIT_API_KEY=your_livekit_api_key
LIVEKIT_API_SECRET=your_livekit_api_secret
LIVEKIT_URL=wss://your-project.livekit.cloud
Step 3: Run the Development Server
bash


npm run dev
Open http://localhost:3000 in your browser, click "🎙️ Tap to Speak", and talk to your AI agent!
Enter fullscreen mode Exit fullscreen mode

Top comments (0)