Puch AI

AI for India, built in‑house.

From 2025 to 2026 we built three products and the infrastructure under them.

01 · What Puch was

An AI assistant inside WhatsApp.

No app, no sign-up. People messaged one number and asked, in their own language, by text, voice note or photo.

Reach

A million people, most of them new to AI.

The product

Real conversations.

Cricket updates in Hindi Fact check of a forwarded message Marathi recipe Voice reply Image generation Match scores Fact check
Beyond consumers

Businesses and governments asked for it too.

Inbound requests

From businesses and public programmes.

AP

Government white-label

"Adugu" for Andhra Pradesh's MEPMA programme.

MCP

Open to developers

Any MCP server could plug into WhatsApp.

02 · How we ran it at scale

Millions of messages, on a free product.

Off-the-shelf APIs were too slow, too fragile and too expensive at that volume, so we built the core pieces ourselves.

Load

Hundreds of thousands of messages a day.

Built in-house

Six things we built in-house.

AI gateway

Our models first, automatic fallback.

→

Search

Our own search and crawl cluster.

→

Media generation

Images and video on our GPUs.

→
03 · AI gateway

One gateway for every model call.

Every answer started with a model call. So we built a gateway: open models on our own GPUs first, model APIs only as a fallback.

How it routed
Appone OpenAI-style call
Gatewayhealth · breaker · weights
Tier 1 · self-hostedopen models on our GPUs
Tier 2 · model APIsfallback only
05 · Media generation

A media studio on our own GPUs.

People asked for pictures, edits, stickers and videos all day. Running our own models kept that affordable.

0
cold starts: every model stayed loaded
06 · Vartalaap

Our app and website, built in stealth.

Here we shipped an assistant you could talk to, and an agent that could use a browser and a phone for you.

Live mode

Talk naturally. Interrupt anytime.

Real recordingExplainer

Browser use

A real browser, live on your phone.

Real recordingExplainer

Mobile use

It used apps on your phone, with your permission.

Real recording on AndroidExplainer

Voiceyou ask
Seelive screen
Planstep by step
Act14 gestures
Askbefore pay or send
07 · Guftagoo

One inbox for people and AI.

An Android app that brought your WhatsApp chats and Puch AI into one place. You signed in with a phone number and a code, and that was it.

Real recording on AndroidExplainer

How it worked

Your WhatsApp chats came with you.

Sign-in with a phone number Puch AI waiting in the chat list Linking WhatsApp with a pairing code WhatsApp chats in the inbox, each with a badge A reply sent through WhatsApp
Phone number+91 by default
SMS code6 digits
Puch AI chatalready waiting
Pairing codetyped into WhatsApp
WhatsApp chatswith a badge

Puch AI built in

Every user got their own Puch AI chat, and the search bar read "Ask Puch AI or Search".

WhatsApp bridge

Linked with a pairing code. Replies still went out through WhatsApp.

Built on Matrix

An open messaging standard, with Element X as the base, so more apps could plug in.

08 · Engineering

By the numbers.

Shipped in Puch

  • WhatsApp assistant
  • Message platform
  • AI gateway
  • Search
  • Media generation
  • Indian-language speech

Shipped in Vartalaap

  • Live mode
  • Phone calls (SIP)
  • Browser use
  • Mobile use

Research

  • Code-mixed speech recognition
  • Personal VAD
  • Neural echo cancellation
  • 4-bit media stack