A long walnut bench in a dark product lab: a family of finished edge-AI devices, printed enclosure halves, a bare board, a GPU workstation and a 3D printer behind, one cyan light line glowing on the home hub

The AI-first product lab

AI that ships with screws included.

Custom models from 150M to 8B parameters, each trained for one niche. The board they run on. The enclosure they ship in. The app and the platform around them. One lab, one thread, from sketch to purchase order.

ModelBoardEnclosureFirmwareAppPlatform

Scroll

A very wide view of one continuous lab bench with six stations: a monitor with training curves, a board under a microscope, a 3D printer with an enclosure half, a soldering station, a phone and tablet on stands, and a small server cabinet, tied together by one cyan light line

One lab, one thread

Six stations. One bench.

  1. 01 · ModelTrained for the job150M to 8B parameters, one niche, scored on the target silicon before anything ships.
  2. 02 · BoardSilicon chosen by the modelModule, radios, antenna placement and the power budget, decided by what has to run.
  3. 03 · EnclosurePrinted by lunchCasings on a Bambu Lab X2D, light guides on an Elegoo Saturn 16K, designed to be moulded later.
  4. 04 · FirmwareSurvives the fieldBoot, provisioning, the inference loop, and an update path that cannot brick a device in a hallway.
  5. 05 · AppPhone, TV, webThe surfaces a customer touches, built by the same people who built the device.
  6. 06 · PlatformDay 400, handledTenancy, connectors, OTA, telemetry and a developer surface. The part nobody demos.

What we actually do

We build the whole animal. Not the skin.

Most "AI products" are a login screen in front of somebody else's datacenter. Ours have a part number. We take one problem all the way down, through the model weights, the PCB, the plastic and the update channel, because that is the only altitude at which local AI is a product and not a demo.

Macro of a system-on-chip with a cyan trace glowing along the board

01 / Models

Small models that know one thing cold

From 150M parameters on a microcontroller to 8B on a box in the building. Trained on your domain, quantised for your silicon, scored before they ship. Yours, with no per-token meter and no deprecation notice.

  • Speech
  • Vision
  • Sensor fusion
  • Local reasoning
A resin-printed optical light guide catching a cyan glow

02 / Hardware

Things with a weight and a BOM

Enclosures, light guides, thermals, antenna placement, the boring bracket that decides whether it survives a real installation. Printed in-house, documented for a factory.

An instrumented test rig with thermal probes on a small device

03 / Electronics

Boards that fit the model, not the other way round

Module and silicon selection driven by what has to run. The radio stack, Wi-Fi, Thread, Matter, Zigbee, Bluetooth, and where the antenna can actually live. Firmware that survives the field.

A compact rack appliance with a cyan status line in a dim server room

04 / Platform

The unglamorous part that decides everything

Provisioning, tenancy, connectors, OTA updates, telemetry, apps for phone and TV, a developer surface. The part nobody demos and everybody needs on day 400. Two of our products already share it.

  • Multi-tenant
  • Hosted or self-hosted
  • Connectors & SDKs
  • OTA & telemetry

Custom models

From 150M to 8B. One niche at a time.

Frontier models are extraordinary and completely unbothered by your problem. A small model trained on the things that can actually happen in your product beats them on the only axis you care about, and fits on a chip you can afford.

150M500M1B3B8B

150M

Fits in the device

Wake word, intent, detection, sensor classification. Runs on the silicon inside the product, offline, in milliseconds.

8B

Fits in the building

Structured, tool-calling reasoning on a box on the premises, for work whose data was never going to be allowed out.

Home automationIoT & industryPhysical securityCyber securityFinanceCompliance

Inside the model work

Cloud AI asks permission. Local AI just answers. One of those is a product feature, and the other is an apology you make to a customer at three in the morning.

The offer, abridged

Three honest numbers, in weeks

No discovery phase that discovers itself. Each stage ends in something you can hold, run or hand to a factory.

2weeks

Discovery

What runs on device, what cannot, which sensors, which model, which silicon, what it costs to make. An architecture and a build plan. Occasionally the advice not to build it.

4weeks

Working software

Inference on the target, the app, the API, the data path. Not a slide deck of a product: the product, early and unflattering, in your hands.

8weeks

Hardware prototype

A working prototype in its enclosure, with the bill of materials and the documentation set a manufacturer needs to quote mass production.

What each stage includes

Start here

Tell us what has to happen in under a second.

Bring the awkward problem, the one that needs an answer before the round trip to a datacenter finishes. Two weeks from now you will know whether it can be built, what it costs, and what it weighs.