What we do
From one conversation to a twin you can ask anything.
How a TwinSetu twin is made, how it answers, how we check it, and the consent it rests on.
How it works
One interview. Every question after it.
Interview
An hour or two of conversation in the person’s own language, recorded with their consent. An AI interviewer leads it, or a human one does, with an AI beside them that transcribes and suggests questions.
- AI-led
- Interviewer-led, AI-assisted
Twin
Built from that transcript and the person’s own particulars, nothing else, so it answers in their terms rather than a typical person’s. Where the conversation never went, it reasons from their values and circumstances and commits to an answer.
Ask
Any new question — a concept, a price, a scenario — to one person’s twin, to a few side by side, or to a whole panel in one run.
- Talk to one
- Ask a few
- Ask a whole panel
People, not profiles
We don’t start from a type. We start from a person.
Two people of the same age, in the same town, on the same income will tell you different things about the month-end, the family WhatsApp group, Friday’s release and who at home really decides. A guess from age, gender and place describes an average Indian, and nobody you need to reach is average.
Anyone can generate a billion plausible Indians. Knowing which answers to trust is the work.
In Indian languages
In the language they speak at home.
People say more in the language they think in, so the interview is spoken, transcribed and asked in theirs. What the twin reads is a checked English translation of that record.
A language joins when its interviews have been piloted with people who speak it and its translations proof-read by someone who reads it.
- English
- हिन्दीHindi
- ଓଡ଼ିଆOdia
India, as a thousand people
Hindi and Odia — the mother tongue of 467 in every 1,000 Indians.
Languages, not people: the ones we can interview in today.
Every other language is still to come.
See the numbers
| Mother tongue | In every thousand | People |
|---|---|---|
| Hindi | 436 | 52,83,47,193 |
| Odia | 31 | 3,75,21,324 |
| English | 0 | 2,59,678 |
| Bengali | 80 | 9,72,37,669 |
| Marathi | 69 | 8,30,26,680 |
| Telugu | 67 | 8,11,27,740 |
| Tamil | 57 | 6,90,26,881 |
| Gujarati | 46 | 5,54,92,554 |
| Urdu | 42 | 5,07,72,631 |
| Kannada | 36 | 4,37,06,512 |
| Malayalam | 29 | 3,48,38,819 |
| Punjabi | 27 | 3,31,24,726 |
| Assamese | 13 | 1,53,11,351 |
| Maithili | 11 | 1,35,83,464 |
| Santali | 6 | 73,68,192 |
| Kashmiri | 6 | 67,97,587 |
| Nepali | 2 | 29,26,168 |
| Sindhi | 2 | 27,72,264 |
| Dogri | 2 | 25,96,767 |
| Konkani | 2 | 22,56,502 |
| Manipuri | 2 | 17,61,079 |
| Bodo | 1 | 14,82,929 |
| Sanskrit | 0 | 24,821 |
| Other languages | 33 | 3,94,91,446 |
- The Census counts Bhojpuri, Rajasthani, Chhattisgarhi and dozens of other mother tongues under Hindi.
- English is rarely anyone’s mother tongue and widely a second language, so it lights no dot here.
- 22
- languages in the Eighth Schedule of the Constitution
- 19,569
- mother tongues written on the Census forms of 2011, before grouping
- 3
- languages we interview in today
Sources: the Constitution of India; Census of India 2011, Paper 1 of 2018: Language (Table C-16).
How we check it
How we know a twin is any good.
Every twin answers questions its person has already answered, and we compare. It has to beat a guess made from age, gender and place alone, and it is held to how consistently the person answers the same questions themselves. The same twin handed a stranger’s interview instead shows how much of each answer comes from the person.
agreement with the person’s own answers
where a twin lands
what the interview is worth
- Age, gender and place — the floor
- A stranger’s interview — the check
- Their own interview — the twin
- The person, asked again — the bar
Every twin comes with its paperwork
Expert observers
Eight specialist readers, from a psychologist to an anthropologist, annotate every interview, or say there is not enough to go on.
Twin report
How closely each twin matched its person, item by item, written up and stored.
Audit trail
Every step from the recording to an answer, kept as a trail you can open.
Consent
Every twin is someone who said yes.
Each twin is one person who sat for an interview, agreed in their own language to what it would become, and can take that agreement back.
- Asked, not scraped
- Built from an interview the person chose to give, never from social feeds, call records or purchase histories.
- Theirs to withdraw
- If they withdraw, their interview and their twin are erased.
- Every step on the record
- Each step their words take is kept as a trail the team can open.
- Not a stereotype
- Every twin is measured against the guess a few demographic labels would make, so a stereotype cannot pass for a person.
Fair questions
Is a twin the person?
No. It is a model of how one person answers, built only from what they told us, which is why every answer it gives is checked against the person themselves.
What happens when a question goes beyond the interview?
The twin reasons from the person’s values, habits and circumstances and commits to an answer. That is exactly why each answer is checked against the floor, the bar and a stranger’s interview.
Does the twin read the person’s own language?
It reads a checked English translation of the transcript. The conversation itself, and the questions put to the person, are in their language.
Who do you want to hear from?
Tell us who you need to understand and what you would ask them. We’ll show you what a panel of twins can tell you.
Bring us your questionor write to hello@twinsetu.ai