01 — What is Lumi

AI,
out of the chatbox.

The Lumi desk agent
Lumi's arm reading a desk
Senses

What is happening on your desk

Lumi knows what you are doing on your desk in real time — paperwork, circuits, or a gadget you just unboxed.

Someone absorbed in folding paper at a sunlit desk
Assists

Without breaking your focus

Lumi knows when you need help and steps in without you asking.

A padlock glowing on a processor chip
Runs

Entirely on your device

Lumi runs completely offline, so nothing you work on ever leaves the desk.

Your AI copilot that guides you without prompting

Live demo
Lumi sees
A half-built sensor board, and its schematic open on the laptop behind.
On the desk
The breadboard. Lumi checks the board against the schematic on screen and flags the wrong row before you power it.

A soldering iron in one hand, a schematic on the screen

Both hands are busy and half the problem is physical. Lumi holds the bench and the file in one thought, so you can ask about the part under your iron without putting it down.

Inspecting
I see a resistor in the wrong row.
A remote worker on a video call with Lumi showing talking cues
Lumi sees
Four people on the call, one agenda in the notebook, two points still unsaid.
On the desk
Lumi's face. Your prompts sit on the desk, not the shared screen. Nobody on the call sees them.

Your next three points, out of frame

Lumi keeps your agenda on its own face where only you can read it, tracks what has been covered, and writes the minutes while you talk.

Talking cue
Ask for feedback · confirm next steps
Printed reports and a notebook on a desk with Lumi reading them
Lumi sees
A printed quarterly, a notebook of margin notes, and the same figures open on screen.
On the desk
The printed report. Lumi lifts the numbers off the page, checks them against the open file, and flags the two that disagree.

Paper and file, in the same conversation

Serious reading still happens on paper, where every assistant loses the thread. Slide a page under the light and Lumi picks up the figures and keeps your handwriting in the record.

Reading
The table on page 3 matches your open file.
Lumi on a tidy desk beside a monitor and keyboard
Lumi sees
A task list open, hands still moving, forty minutes unbroken. Nothing worth interrupting.
On the desk
The monitor. Lumi watches the work, not you: which document, which task, which stage.

Six hours at a desk, and something watching the work

Most of the day is not a task you can name. Lumi holds the thread of what you are doing and speaks up at the pause, not the moment it has something to say.

Standing by
You are mid-sentence. I will hold.
Privacy

Nothing leaves
your desk.

Most AI assistants send your screen to someone else's server. Lumi's models run on the hardware in front of you.

  1. 01
    Runs on the device
    Seeing, listening and deciding all happen on Lumi’s own chip.
  2. 02
    Offline by design
    Pull the network cable and Lumi keeps working exactly the same.
  3. 03
    Nothing leaves by default
    Your desk is never uploaded, reviewed, or trained on. If anything should ever leave, you say so.
Beyond work

Off the clock,
still by your side.

The desk is where we start, because that is where the hours are. But an assistant with eyes and a body can know your space, not just your files.

What makes an assistant personal was never in your documents. It is in your room, your habits, the things within reach.

A hand in the kitchen, Lumi watching the pan
In the kitchen

It watches the pan with you and speaks up before the garlic turns.

Lumi keeping the house in order
Around the house

It knows where things belong, and what moved since yesterday.

Lumi turning to meet someone walking in
For the company

It turns to meet you when you walk in — presence you can feel, not another app to open.

About us

Empower human intelligence.

Today's AI agents live inside the screen. We use multimodal AI to bring them out — into the room where your real work happens. That is why our mission is to empower human intelligence, not replace it. Lumi is our first step in extending AI's capability to the real world.

The team

Frontier embodied AI research × mass-production autonomy

Benjamin Li
Benjamin Li
Co-founder & CEO

BS and MS in Mechanical Engineering, UC Berkeley. Led autonomous driving perception from zero to one at Xpeng and ZEEKR.

UC BerkeleyXpengZEEKR
Expertise
Multimodal perception · On-device inference · Product definition
Tom Huang
Tom Huang
Co-founder & CTO

Research Scientist, Boston Dynamics. Core author of an ICRA 2024 best paper. Works on robot learning and spatial reasoning — how a machine models the space it works in.

Boston DynamicsTUMUniversity of Freiburg
Expertise
3D scene understanding · Embodied agents · Robot learning
John Zeng
John Zeng
Co-founder, Head of Software

Embedded systems and multimodal perception. Took on-device AI products from prototype to manufacture, and built production perception systems alongside Benjamin.

Zhejiang UniversityZEEKRvbot
Expertise
Edge inference · Sensor fusion · Hardware-software integration
Early access

Meet Lumi first.

Leave your email. One click, no account.