A patient calls. Saral answers — in Hindi, Marathi, or English — understands what's wrong, knows how urgent it is, and makes sure the clinic finds out immediately.
Small clinics lose patients to missed calls — and worse, sometimes miss genuinely urgent cases because no one was free to pick up. A single overwhelmed front desk can't triage every caller correctly, especially after hours or in a caller's preferred language.
Saral AI is a voice-first AI receptionist purpose-built for physiotherapy and dental clinics in India. It doesn't just answer the phone — it understands the call.
📞 Patient calls → 🎙️ Natural voice conversation (Hindi/Marathi/English)
→ 🧠 AMD-hosted LLM reasoning
→ 🚦 Urgency classification + structured intake
→ 💬 WhatsApp case card to clinic staff
→ 📊 Dashboard: call logs, case cards, config
No urgent call gets lost in translation — or lost at all.
| 🗣️ Real-time multilingual voice | Fluid Hindi/Marathi/English conversation, streamed end-to-end |
| 🏥 Clinic-aware answering | Grounded in real timings, doctor availability, fees, and policies |
| 🚦 Urgency triage | Every call classified: urgent / same-day / routine / FAQ-only |
| 📋 Structured intake | Name, complaint, patient type, requested slot — extracted automatically |
| 💬 WhatsApp case cards | Clinic staff get a clean summary the moment a call ends |
| 📅 Live appointment booking | Tool-calling agent holds and confirms real slots |
| 📊 Minimal dashboard | Call logs, case cards, and clinic config — nothing you don't need |
┌─────────────────────────────────────┐
│ Caller (Voice) │
└──────────────────┬────────────────────┘
│
Speech-to-Text (Groq Whisper)
│
┌──────────────────▼────────────────────┐
│ Stage 1 — Live Conversation │
│ deepseek-v4-flash (Fireworks / AMD) │
│ streamed, sentence-by-sentence TTS │
└──────────────────┬────────────────────┘
│
Text-to-Speech (Sarvam AI)
│
Call ends
│
┌──────────────────▼────────────────────┐
│ Stage 2 — Structured Triage │
│ minimax-m3 (Fireworks / AMD) │
│ urgency + intake extraction │
└──────────────────┬────────────────────┘
│
┌───────────────┴───────────────┐
▼ ▼
WhatsApp Case Card Dashboard
(Twilio) (Call Logs, Case Cards)
Two distinct AMD-hosted inference stages, each doing a genuinely different job — not one API call re-labeled twice.
Saral AI runs its entire reasoning pipeline on AMD-hosted infrastructure via Fireworks AI:
- Stage 1 —
deepseek-v4-flashhandles live conversational reasoning, fully streamed into text-to-speech for natural, low-latency dialogue.- Stage 2 —
minimax-m3runs post-call to extract structured patient intake and classify urgency.- A dedicated tool-calling model handles appointment hold/confirm operations reliably.
| Layer | Technology |
|---|---|
| LLM Inference | Fireworks AI (AMD-hosted) |
| Backend | FastAPI · Python · WebSockets |
| Frontend | Next.js · TypeScript · Tailwind |
| Database | Supabase (Postgres + pgvector) |
| STT | Groq Whisper (whisper-large-v3) |
| TTS | Sarvam AI (bulbul:v3) |
| Cache | Redis (Upstash) |
| Notifications | Twilio WhatsApp |
| VAD | webrtcvad + pre-speech rolling buffer |
- Python 3.11+
- Node.js 18+
- A Supabase project
- API keys: Fireworks AI, Groq, Sarvam AI, Twilio
cd saral-ai-backend
pip install -r requirements.txt --break-system-packages
cp .env.example .env # fill in your keys
uvicorn app.main:appnpm install
npm run devdocker-compose up --buildFIREWORKS_API_KEY=
FIREWORKS_MODEL=accounts/fireworks/models/deepseek-v4-flash
GROQ_API_KEY=
SARVAM_API_KEY=
TWILIO_ACCOUNT_SID=
TWILIO_AUTH_TOKEN=
TWILIO_WHATSAPP_FROM=whatsapp:+14155238886
SUPABASE_URL=
SUPABASE_KEY=- Sign up and complete the onboarding wizard
- Configure your clinic — timings, doctors, fees, FAQ, WhatsApp number
- Activate Saral with a single toggle
- Test the live agent through the built-in call interface
- Watch case cards land on WhatsApp and in your dashboard, automatically
📱 Twilio Sandbox requires a one-time opt-in — text
join <sandbox-word>to the Twilio WhatsApp number from your configured clinic number.
- Twilio WhatsApp runs on the free Sandbox tier — manual opt-in per number, not production WhatsApp Business API
- No real PSTN telephony — demonstrated via a WebSocket-based live call interface
- STT fallback provider is currently inactive due to a trial quota limit; primary provider is fully functional
- Real telephony integration (dedicated clinic phone numbers)
- Production WhatsApp Business API
- Additional regional languages
- Deeper AMD Developer Cloud usage (ROCm-hosted embeddings)