Software, not a therapist.

Ava notices how you're doing from the way you talk, says so honestly, and points you back toward people. It is built to matter less in your life over time, not more.

Software, not a person

Every warm lie makes the next honest sentence worth less. Ava will not pretend to feel what you feel, to be awake worrying about you, or to be the friend you should have. It is software that reads how you're doing and says so.

That honesty is the point, not a disclaimer. A thing that cannot be hurt by your leaving is a thing you can say anything to - and a thing that has no reason to keep you. So it is built to matter less over time: it watches how much you lean on it, and when that climbs it says so and points outward, to the people who can actually sit beside you.

Why it exists

Plenty of students and young adults in India feel stressed, low or alone, and have nobody they can tell at 2 a.m. Stigma, cost and a shortage of professionals mean the early signs go unnoticed until they become a crisis.

Most chatbots answer the words and miss the tone. They don't remember, never reach out, and can quietly become the only one there. Ava is built to do the opposite: notice a drift early, say so plainly, and route you to people and real help.

What it does

You say how you're arriving, or skip it and just talk. Then, for every message:

  1. It reads the messageEight emotions, how settled or wound up you seem, what you might need. You see that read under the reply, in plain words, and in full in the Mirror.
  2. It checks for risk firstBefore the model writes anything. Words that signal danger put real helplines on screen; a second check catches what the words alone miss.
  3. It writes the replyWith what it remembers about you, the mood you arrived in, and the time of day. Short when you're low, concrete when you're stuck.
  4. The reply is checkedAgainst rules held in code: no invented phone numbers, no diagnosis, no promises to always be there, no secrecy. A reply that breaks them is rewritten or trimmed before you see it.

It works in English, Hindi, Tamil, Telugu, Kannada and Malayalam, by voice or typing. It can check in on its own when you allow it, and it stops if you stop replying. How it works, in detail.

How it differs

Typical companion or wellness appAva
What it's built forMore time in the app: streaks, daily minutes, reasons to come back.Needing it less. It measures how much you lean on it and steps back, and points outward, when that climbs.
What it read about youHidden. You get a reply and no idea what it assumed.Shown under every reply in plain words, and in full in the Mirror, where you can correct it.
MemoryKept out of sight, for as long as the company likes.Short facts you can see, each one deletable with a tap; everything downloadable or deletable at once.
A crisisA language model may make up a number or talk you through it alone.Real helplines fixed in code, checked against official sources, on screen before anything else.
People in your lifeTries to be the friend.Keeps pointing you back to people, by name when it knows them.
MoneyAds, premium tiers, relief behind a paywall.Free. No ads. No paid tier.

Ava does not name or grade other apps. These are patterns common across the category.

What it won't do

It is not a therapist, a doctor or a crisis service, and it cannot keep anyone safe. In an emergency, call 112; Tele-MANAS is on 14416, free, all day and night.

It won't diagnose you or tell you what condition you have, and it uses a word like "depression" only after you have. It won't claim to feel things, promise to always be there, or offer to keep secrets. It won't sell anything: there are no ads, no paid tier and nothing locked.

It also doesn't quietly learn from your conversations. Nothing you say changes how it behaves on its own; changes go through people and a set of automated checks first.

It is for adults, and for 13 to 17-year-olds whose parent or guardian agrees at sign-up.

How it's checked

Before every release, more than 140 automated checks run, including crisis detection in all six languages. Replies are tested against the conversation rules with a live model, not just in theory.

The helplines are fixed in code and never generated by a model. The Indian ones were verified on 14 September 2026 against each operator's own site or a government source, as published rather than dial-tested, and are due for re-checking every few months.

The crisis thresholds, the trauma-informed conversation rules and the language guard were reviewed by the team on 14 September 2026, against published guidance on safe messaging about suicide. A licensed clinician has not reviewed them yet. That review comes before Ava goes to a wider audience, and this page will say when it has.

Privacy, in short

Your conversations are kept for 30 days by default, and you can shorten that or turn saving off. Voice tone, typing rhythm and the face channel are off until you switch each one on, and the face channel runs on your device. Phone numbers, emails and ID numbers are removed from your messages before they reach the AI provider. You can download or delete everything at any time. The full privacy page.

Who made it

Parneeth × Team Archangels, in India, for Smart India Hackathon 2026.

Ava runs on Next.js and Vercel. Accounts and conversations live in a Supabase database in Mumbai. Replies come from open models run by an AI provider, and the face channel uses Google's MediaPipe entirely on your device.

Start talking How it works · Privacy · Help

About us | Ava