albusforge.ai
About us

Physical AI, built from a sentence

One question in. A working device out, and an AI cloud that closes the loop.

Why we’re building this

Physical AI should be something anyone can build.

Making a device that senses and acts still means choosing parts, wiring, writing firmware and building cloud plumbing, and every step needs expertise. We think a plain sentence should be enough.

A gadget that works is a weekend project. A gadget whose data turns into answers, and then into action, is one worth keeping. So we don’t stop at a dashboard: every device we generate closes the loop.

It also has to be trustworthy. That means real parts, numbers that come from the data, and a person in charge of every action.

What we build

From a plain-language request to a working device

  1. Ask01

    A plain-language request

    One question in: “keep my greenhouse soil moist.”

  2. Parts02

    A real parts cart

    Real components from a curated part registry.

  3. Enclosure03

    A 3D-printable enclosure

    Generated around those exact parts.

  4. Firmware04

    Working firmware

    Only the app layer is generated.

  5. Kit05

    A kit

    Parts, firmware and enclosure ship as a kit.

  6. Live06

    Live in the cloud

    The device connects to an AI cloud platform that watches its data, reasons about it and acts back on the device.

The closed loop

Sense → Detect → Reason → Act → Confirm

  1. 1 · Sense

    Sense

    Telemetry arrives already typed: the part registry declares each channel’s unit, range and meaning before the first packet.

  2. 2 · Detect

    Detect

    Statistics run on every data window — learned baselines, seasonality and drift — to find anomalies cheaply, without a model.

  3. 3 · Reason

    Reason

    A small model names and triages each detection; a frontier model answers questions through typed queries.

  4. 4 · Act

    Act

    The platform sends commands back to the device — thresholds, setpoints, schedules — or opens a work order.

  5. 5 · Confirm

    Confirm

    The next telemetry shows whether the action worked, and every confirmed outcome improves the baselines.

The loop starts on the device: reflex rules keep working offline, and the cloud adds judgment on top.

Our principles

What we hold ourselves to

  • One curated part registry drives everything.

    One data object, the Part Definition, threads through every service. No service hard-codes knowledge about a specific part.

  • Every number comes from the data, never from the model.

    Models answer questions through typed queries against your telemetry.

  • A person or a policy confirms every action.

    Actions a model proposes need a person, or a policy, to confirm them.

  • Own your data.

  • Security patches.

    How devices are patched and how your data is handled is written up in our Security pledge.

Team

Team Albus

Sukrit Dasgupta

Solo hacker, Team Albus

Being built at MIT, for the Battle of the Coasts hackathon, in the Deep Tech / Physical AI track.

Get in touch

Describe the device you want.

Save your build. Own your data.

Start a build →Create an account