Nandi the bull resting in a field beside a small voice assistant on a wooden stool, sound waves rising from the device toward his ear, a red sun behind them.

Nandi listens, in your language.

A small voice assistant you can talk to in Hindi, Tamil, Bengali and eight more languages. It starts answering in about a second and a half, remembers what you tell it, and changes from a dashboard with no re-flashing.

Watch the demo See how it works
Event logo 240 × 96 px

Built for the KAYA Buildathon, AZMTH x IIT (BHU) Varanasi

Voice assistants were not made for most of India.

What gets in the way What Nandi does

They work best in English.

Nandi hears and answers in 11 languages, and replies in the language you spoke.

You wait before they answer.

Nandi starts speaking about 1.5 seconds after you stop.

One small change means re-flashing the board.

Voice, language, volume and personality change live from a dashboard.

Every gadget works alone.

Every Nandi and every sensor in the house can reach the others.

How it works

  1. Wake it

    Say “Nandi”, or press the button. The wake word is heard on the device itself.

  2. It listens

    The device streams your voice over Wi-Fi to the Nandi server.

  3. It thinks

    Sarvam AI turns your speech into text. Nandi’s agent answers, and searches the web, checks the news or weather, or reads a sensor when it needs to.

  4. It answers

    Sarvam AI speaks the reply. It plays through the speaker while the bull on the screen acts it out.

How the parts connect

Nandi device

  • Microphone and wake word
  • Speaker
  • Bull screen
your voice the answer

Nandi server

  • Agent and tools
  • Memory
  • Recordings and settings
speech, text text, speech

Sarvam AI

  • Speech to text
  • Language model
  • Text to speech

Also connected to the server:

The dashboard Other Nandis Sensors

It does more than answer questions.

Remembers you
Tell it your father’s name once. Weeks later it says “you told me…”, in your language.
Reminders and timers
“Remind me at six to call Amma.” Set by voice, spoken out loud on time.
Talks between rooms
“Tell the kitchen dinner is ready.” Nandi says it on the kitchen device.
Knows today’s news
Live web search, news and weather through a private search engine. When an answer sounds unsure, Nandi checks again.
Changes by voice
“Speak slower.” “Answer in Tamil.” Its settings change as you ask.
Speaks without internet
Messages like “Wi-Fi is not working” are stored on the device, so it can still tell you what is wrong.
A grandmother, a father and a young girl in their living room smile at a Nandi device speaking from a wall shelf.
People from different parts of India speak toward one small Nandi device, their voices drawn as coloured ribbons flowing into it.

Eleven languages. It answers in yours.

Hearing, thinking and speaking all come from Sarvam AI, which is built for Indian languages. Ask in Hindi and Nandi answers in Hindi. Ask the next question in English and it follows you.

  • Hindiनमस्ते! मैं नंदी हूँ, आपकी सेवा में।
  • EnglishHello! I am Nandi, happy to help.
  • Bengaliনমস্কার! আমি নন্দী, আপনার সাহায্যে আছি।
  • Gujaratiનમસ્તે! હું નંદી છું, તમારી મદદ માટે તૈયાર.
  • Kannadaನಮಸ್ಕಾರ! ನಾನು ನಂದಿ, ನಿಮ್ಮ ಸಹಾಯಕ್ಕೆ ಸಿದ್ಧ.
  • Malayalamനമസ്കാരം! ഞാൻ നന്ദി, നിങ്ങളെ സഹായിക്കാൻ തയ്യാർ.
  • Marathiनमस्कार! मी नंदी, तुमच्या मदतीला तयार.
  • Odiaନମସ୍କାର! ମୁଁ ନନ୍ଦୀ, ଆପଣଙ୍କ ସାହାଯ୍ୟ ପାଇଁ ପ୍ରସ୍ତୁତ।
  • Punjabiਸਤ ਸ੍ਰੀ ਅਕਾਲ! ਮੈਂ ਨੰਦੀ ਹਾਂ, ਤੁਹਾਡੀ ਮਦਦ ਲਈ ਤਿਆਰ।
  • Tamilவணக்கம்! நான் நந்தி, உங்களுக்கு உதவ தயார்.
  • Teluguనమస్కారం! నేను నంది, మీకు సహాయం చేయడానికి సిద్ధం.

Meet the bull on the screen.

A 1.3-inch screen shows Nandi’s world: a hut, grass, and a sun and moon that follow the real clock. The bull raises his ears when he listens, shows a thought bubble while he thinks and moves his mouth as he speaks. Ask him to dance, eat or sleep and he does it.

The bull on Nandi's screen standing in the grass beside his hut, under the sun and a cloud, drawn in dots.
This is the device’s own animation engine, drawn dot for dot.

27 animations in all, drawn on the device at 15 frames a second.

One Nandi, or one in every room.

Every device runs the same firmware and signs itself in. Settings apply to the whole house or to a single device. Devices can speak on each other, and they can read each other directly over the home Wi-Fi.

Sensors join in

Any small ESP32 can report readings with a few lines of code. Our health sensor measures pulse with a MAX30102. Ask “what is the heart rate?” and Nandi reads it out.

A two-storey home cut open to show four rooms, each with a Nandi device, joined by dotted lines. In the bedroom an older man wears a pulse sensor on his finger.
Two hands hold a phone showing simple controls, with dotted lines running to Nandi devices in a kitchen, on a bedside table and on a study desk.

Change it from a dashboard. Never re-flash.

  • See every conversation, with the recording of the question and of the answer.
  • Make any device speak: type a sentence and press Enter.
  • Tune voice, language, answer length and personality for the house or for one device.
  • Add or forget what Nandi remembers, with undo.
  • Works on a phone and on a projector.
Dashboard screenshot 1600 × 1000 px, PNG or WebP

Updates that undo themselves if something breaks.

  1. Publish

    Upload the new firmware in the dashboard. The server keeps a fingerprint of the file, and each device checks it before installing.

  2. Try it on one

    Pilot devices get it first. Each one asks out loud before it installs, and never during quiet hours.

  3. Release to all

    When the pilots report back, one click offers it to every other device.

If a new version cannot reach the server within three minutes, the device goes back to the old version by itself.

Built from common, low-cost parts.

ESP32-S3 (N16R8)
The brain. 16 MB flash, 8 MB memory, Wi-Fi.
INMP441 microphone
A digital microphone that hears you.
MAX98357A amplifier and speaker
Plays the answer.
1.3-inch OLED screen
128 × 64 dots, home of the bull.
One button
Press to talk, when you do not want to say “Nandi”.
A vintage-style exploded drawing of the Nandi device showing its shell, screen, circuit board, microphone, amplifier and speaker.
Photo of the device 1200 × 900 px, JPG or WebP
Photo of the board 1200 × 900 px, JPG or WebP

Measured, not guessed.

11
languages it hears and speaks.
1.5 s
median time from the end of your question to the start of the answer, over 545 real questions.
97.4%
of wake words caught in testing, with 2.4% false wake-ups, from a 14 KB model on the device.
27
bull animations, from sleeping to celebrating.

What comes next

  • More languages and voices as Sarvam AI adds them.
  • Smart-home control: switch lights and fans by voice.
  • A battery build you can carry from room to room.
  • More of the listening and thinking on the device itself.
  • An open kit, so other students can build their own Nandi and add devices to the family.

The team

Team Novadian, IERT Prayagraj
Institute of Engineering and Rural Technology

  • Anant Tiwari Team leader
  • Ashutosh Kushwaha Team member
  • Prateek Singh Team member

ananttiwari2406@gmail.com
github.com/anant-dev-ok

Team logo240 × 96 px Institute logo240 × 96 px Event logo240 × 96 px

Or watch it on YouTube.