Menu

Work · Our own product

A voice assistant we made end to end, from the board to the launch.

We designed the circuit board, wrote the firmware, built the AI server and the dashboard, and made the launch website. One team, every layer.

A 3D render of the Nandi v1 circuit board with its wireless module, USB-C port and TALK button.
Indian languages
11
Median voice reply
1.5 s
Automated tests
909
Custom 2-layer board
70 × 60 mm

Nandi is named after the bull who sits facing the shrine and listens. Our Nandi listens too: you speak to it in your own language, and it answers out loud in about a second and a half.

We made every part of it ourselves. That is why it is the first story on this site.

The brief

We wanted a voice assistant for Indian homes: one that understands many Indian languages, answers fast, and runs on a small, low-cost board. So we built one, and we keep making it better.

The board

The top of the Nandi v1 board: the wireless module at the top, speaker and battery connectors at the sides, and USB-C, RESET and TALK at the bottom.
fig. 01 — Nandi v1 main board, 70 × 60 mm, designed in-house

The main board is ours, from the schematic to the factory files:

  • Brain: a wireless module with Wi-Fi, Bluetooth and plenty of memory for sound.
  • Sound: a digital microphone, and an amplifier for the speaker.
  • Power: USB-C, battery charging, and an automatic switch between USB and battery.
  • Screen: a socket for a small OLED display.

The board passes its electrical rule check and its design rule check with zero errors, and the files are ready for a factory to make and assemble. We even describe the board in code: one file holds every part and connection, and scripts turn it into the schematic, the layout and the factory files.

An exploded drawing of the Nandi device: case, screen, main board, microphone, amplifier, speaker and back cover.
fig. 02 — How the parts fit together

The firmware

The board runs our own firmware. A 14 KB wake-word model listens for “Nandi”. In our tests it caught the word 97.4% of the time, with 2.4% wrong wake-ups. Firmware updates arrive over Wi-Fi, and each new version goes to a pilot device before the rest.

On the screen lives a small bull with 27 animations. It blinks, listens with its ears up, and even runs into its hut when it rains.

Nandi's bull character, drawn in dots on the small OLED screen, standing by its hut and listening.
fig. 03 — The bull on Nandi's OLED screen, listening

The brain

When you speak, the board sends your voice to our server. The server uses speech and language AI to understand you and to answer out loud. It supports 11 Indian languages. Over 545 real questions, the median reply came in 1.5 seconds.

The server has about 909 automated tests, so we can change it without fear.

The Nandi control room: live conversations in several Indian languages, the list of devices and a weekly chart.
fig. 04 — The control room: every conversation, every device

The control room shows every conversation with its recording, makes any device speak, tunes voice and language for one device or the whole fleet, and rolls out firmware safely.

The launch

We also made Nandi’s launch website: plain, fast HTML with a set of illustrations in one style.

An illustration of people from across India speaking different languages to a Nandi device.
fig. 05 — From the Nandi launch site

What it shows

The board, the firmware, the cloud, the dashboard and the marketing were planned by one team, so they fit together the first time. That is how we work on every project.

Services

Next project Meetings platform →

Got a new idea? Or an old product?

A new idea, or an old product to upgrade?Tell us what you need →