Voice
Primary. Tap-to-talk in Hindi or English.
How it works
From a sentence
to a system
that remembers.
Voice in, structured care out. Seven steps inside SAHAI — and the architecture behind them.
01step
Speak
The elder taps a big button and says what they feel. "My head hurts." "Remind me to take my BP medicine at 8." Hindi, English, or both.
02step
Understand
Speech becomes text. AI extracts intent, symptom, medicine, time. If it's a critical input, SAHAI confirms — once, gently — before saving.
03step
Append
The event is timestamped and stored. Raw input + structured output, side by side. Nothing is overwritten. Ever.
04step
Schedule
Reminders go into the reliability service — a deterministic queue separate from AI. They will fire on time even if everything else is slow.
05step
Remind
At the right moment, SAHAI speaks: "It's time for your BP medicine." Three big buttons appear: TAKEN, SKIP, REMIND IN 10.
06step
Escalate, only if needed
If a critical medicine is missed, family gets a push. Then a phone call. Then an SMS. Until someone confirms that the elder is okay.
07step
Remember
Confirmed inputs and repeated patterns flow into long-term structured memory — visible in the family dashboard, exportable for the doctor.
Channels
Wherever the elder is,SAHAI meets them.
Text + voice messages, same processing.
App UI
Big-button fallback when voice fails.
IVR call
Reminder calls with keypad input.
Architecture
Five services.One quiet boundarybetween AI and truth.
AI services interpret. They never directly write to the database. The Data Service is the single source of truth — and the Reliability Service runs reminders without ever asking the AI for permission.
Service 01
API Gateway
The entry point. Authenticates, routes, rate-limits.
Service 02
Conversation Service
AI layer only — no database access. Speech, intent, response.
Service 03
Data Service
The single owner of the database. Append-only, audited.
Service 04
Reliability Service
Reminders, escalation, retries. Deterministic by design.
Service 05
Notification Service
Push, voice call, SMS. The chain that must always work.
Tech stack
- Backend — Django / FastAPI
- Database — PostgreSQL
- Queue — Redis (BullMQ)
- Mobile — React Native
- AI — Claude / OpenAI
Offline support
- Inputs stored locally, retried when online
- Reminders fire from the local queue
- SOS works through call + SMS without internet