Skip to main content

Status: live. Delhi --:-- IST · agents running · now buildingLumora

← Back to Work

NamasteParent: AI phone check-ins with elderly parents plus WhatsApp family updates

Live, in development
Role
Architect, full-stack builder and reviewer
Timeline
2026 (Live, in development)
Status
Live, in development · No paying users yet
Stack
Piopiy/TeleCMI Telephony · Gemini 3.1 Flash Live · FastAPI · Next.js · Supabase (JWT + RLS) · Google Cloud Run

Problem

Adults living in another city or country want to check on aging parents back home every day, but time zones and busy workdays get in the way. Meanwhile, elderly parents don't want to install or navigate a health app—they just know how to answer a regular phone call.

Get in touch

What I built

An automated phone check-in service where parents simply pick up a normal phone call from a warm AI voice powered by Gemini 3.1 Flash Live. After the call, the system summarizes how they are feeling, whether they took their medicines, and sends a structured update to the family on WhatsApp.

Get in touch

Architecture diagram

The elderly parent's regular phone connects by PSTN call to the Piopiy/TeleCMI telephony gateway, which bridges bidirectional audio to FastAPI on Cloud Run running a Gemini 3.1 Flash Live audio session. After the call, FastAPI writes to Supabase PostgreSQL, where JWT auth and row-level security isolate each family's data, and the post-call summary flows to the family's WhatsApp update and Next.js dashboard.

A regular PSTN call bridges the parent's phone to a Gemini Live session on Cloud Run; post-call summaries flow from Supabase to the family on WhatsApp and the Next.js dashboard.
Get in touch

Key decisions

  1. 01

    Zero-app PSTN phone call for the parent, WhatsApp for the children

    Why: Kept the parent interface strictly to an incoming regular phone call via Piopiy/TeleCMI and delivered summaries via WhatsApp.

    Trade-off: Higher per-minute telephony cost than an in-app VoIP call, but zero onboarding friction for senior citizens.

  2. 02

    Gemini 3.1 Flash Live for low-latency, interruptible voice

    Why: Used real-time streaming audio (Gemini 3.1 Flash Live) instead of a chained STT → LLM → TTS pipeline so the agent listens while speaking rather than doing "wait-for-your-turn robot talk".

    Trade-off: Requires persistent WebSocket audio bridging on Google Cloud Run, but makes the conversation feel natural to an older listener.

  3. 03

    Strict Row-Level Security (RLS) in Supabase for family health notes

    Why: Enforced JWT-scoped Row-Level Security policies directly in PostgreSQL so family check-in logs are isolated at the database layer.

    Trade-off: More upfront SQL policy testing, but prevents cross-tenant data leaks on sensitive family check-ins.

Get in touch

Result

  • [PLACEHOLDER]

    [ADD METRIC: check-in calls completed]

  • [PLACEHOLDER]

    [ADD METRIC: average call latency / completion rate]

  • [PLACEHOLDER]

    [ADD METRIC: WhatsApp summary delivery time]

Get in touch

Demo

[ADD DEMO VIDEO OR SAMPLE CHECK-IN CALL RECORDING]

Get in touch

Want to build a real-time voice product on Gemini Live or custom telephony?

Get in touch