New
A new name and a new face. Hey Saathi is bigger, softer and kinder, and both screens below are captured from the live build this week.
Try it
It runs in a browser, no install and no account. Open heysaathi.vercel.app on a phone, pick a language, and talk to it the way a parent would.
Venture · 02
Hey Saathi
A voice companion for your parents. It speaks their language, listens without hurry, and never asks them to type.
Spec
- Status
- Live beta
- Try it
- heysaathi.vercel.app
- Languages
- Malayalam, Telugu, Hindi, English
- Stack
- Sarvam speech recognition, Claude
Build
Live beta
0 to 1
Next.js
Sarvam ASR
Claude
Malayalam
Telugu
Hindi
Voice first
Accessibility
01
Millions of elderly Indians live alone while their children work abroad. The phone in their hand assumes young eyes, fast thumbs and English. Video calls happen on Sunday. The other six days are quiet.
Saathi is a companion that talks. In Malayalam, in Telugu, about the old films they love, the news, their medicines, their day. No login, no password, no keyboard. The son in New Jersey sets it up with a pairing code, and the mother in Kochi just speaks.
02
The Flowone call, end to end
01
Family Sets It Up
A son in New Jersey picks the language and reads out a six-character code. The elder never onboards.
02
One Button
No menu, no login, no notification tray. The screen holds a single thing to press.
03
She Speaks
In Malayalam, Telugu or Hindi, at her own pace, with the hesitations real speech has.
04
Intent, Not Transcript
The measure is whether the words that matter survived, not whether every word was spelled right.
05
It Answers Aloud
Spoken back in the same language, slowly, with the type large enough to read without glasses.
06
The Family Sees Enough
That she is using it, not what she said. Company is not surveillance.
She has never used an app. She has used a telephone for sixty years.The person every decision is measured against
03
What I Sweat Overthe details nobody sees
01
Voice First, Screen Second
The whole interface is a conversation. The screen exists to reassure, not to navigate. Anything that needs a tap is a large, obvious button with one job.
02
Understanding Grandma, Not Transcribing Her
Speech recognition for elderly regional voices is hard, so I benchmarked five providers on real elderly Kerala speech. The metric that matters is not word error rate, it is whether the system caught medicine, temple and grandson. Keyword recall over transcription vanity.
03
Setup Happens an Ocean Away
The elder never onboards. Three screens: pick a language, read a pairing code to your child on the phone, done. The family configures everything else remotely.
04
Type You Can Read from Across the Room
Every screen is designed for seventy year old eyes. If a line of text cannot be read at arm’s length without glasses, it does not ship.
04
Screenstype you can read at arm’s length
05
The Deliberate Cutwhat I chose not to build
Chose
Voice first, one button
over a chat interface with a keyboard
Typing Malayalam on a phone keyboard is the barrier, not the conversation. Every character of input the product asks for is a reason to put the phone down. One button removes the entire class of problem.
Chose
Keyword recall as the metric
over word error rate
Word error rate punishes a harmless misspelling as hard as a lost word. What actually matters is whether “chest pain” survives the transcription. I benchmarked five providers on elderly Kerala voices and ranked them on recall of the words that carry meaning.
Chose
Setup by family, from another continent
over the elder onboarding herself
The person who buys this is not the person who uses it. Making the buyer do the configuration removes the exact step where an eighty-year-old would give up, and it turns a support burden into a shared moment.
06
What Isn’t Built Yetthe honest list
Unpublished
Retention numbers exist in the beta but are not public yet. A companion is judged on the second week, not the launch, and that data is still thin.
Deliberately absent
No medical advice, ever, in any language. The product says it does not know rather than guess about a symptom.
Not yet
No offline mode. A dropped connection currently ends the conversation instead of degrading gracefully.
Open question
How much the family should see. Today it is presence only, and that line has not been tested with enough households.
07
Why It Is Built This Waythe person decides
Every decision on this page traces back to one person, and she is not a user in the way software usually means it. She has never installed an app. She has used a telephone for sixty years, which is why the product behaves like a telephone: you press one thing, you talk, it talks back. There is no navigation to learn because there is nothing to navigate.
That constraint decides the rest. It is why the metric is keyword recall and not word error rate, because a companion that mishears the spelling of a temple is fine and one that loses the words chest and pain is not. It is why the family does the setup from another continent, because the person who buys this is not the person who uses it, and the configuration step is exactly where an eighty-year-old would put the phone down. It is why the family sees presence and not transcripts, because company and surveillance are different products and only one of them is worth building.
The honest summary is that almost nothing here is a technical achievement. The speech recognition is bought, the model is bought. What is mine is the set of refusals: no keyboard, no login, no medical advice, no reading her conversations. A product for someone who cannot ask for changes has to be designed by subtraction, because she will not tell you what to remove. She will just stop picking it up.
A week where she talked to it on six days, not one.The north star
Know someone who needs this?
Try It with Your Parents.
It is free while in beta. I want to hear what confused them, that is the whole roadmap.
heysaathi.vercel.app