Physical systems
Hardware that can not act on a guess.
Praetu’s discipline did not begin with chatbots. It began with a harder question: how do you let a learned model take part in a physical action, one with real, immediate consequences for a person, without ever letting it act on an unverified belief? We are applying the answer to AI-enabled hardware for human safety.
The physical world is the strictest auditor.
Software agents fail and someone reviews a log. Physical systems fail and someone gets hurt. Any machine that senses the world, infers what is happening, and acts on a person’s behalf holds two obligations at once: it must be partly learned, because the situations that matter most can't be fully enumerated in advance, and it must be fully accountable, because a safety-critical actuator must never fire on a guess or a lie.
Most AI hardware resolves that tension by trusting the model. We resolve it the other way.
Learned inference proposes. Certified machinery disposes.
Read the world
Redundant sensing, engineered for graceful failure: the system is designed to keep its picture of the world when individual channels degrade or die.
Anchor on the known
The physics that is known for certain is written as verifiable code; learned models cover only the gap: predicting what can't be derived. The model refines the picture. It never owns it.
The certified floor
Between inference and actuation sits a deterministic safety floor with veto power, certified exhaustively over its declared domain, never learned, never argued with. The same boundary architecture proven on our agent benches, with higher stakes.
Never on an unverified claim
Actuation requires grounded conditions, keeps a human confirmation at the seam whenever time allows, and records every request, refusal, and adjustment in an audit trail.
The actuator can not fire on an ungrounded claim: by construction, not by policy.
This is the same property our published agent record demonstrates (completion computed rather than claimed, variance confined to the audit trail) carried into systems where the cost of a false “done” is not a bad ticket but a physical act. The software business and the hardware programs share one discipline; the benches differ.
In private development, bench-first.
The physical programs run the way everything here runs: as declared benches, with predictions registered before results, kill criteria written in advance, and failures recorded at the same prominence as successes. They advance on evidence, not announcements, which is why this page describes a discipline and not a product.
We are not naming domains, publishing specifications, or taking orders. We are open to conversations with serious institutional partners (research, procurement, and program organizations working on AI-enabled safety hardware) under appropriate agreements.
Inquiries
Start a conversation
Institutional and program inquiries only. Tell us who you are and the problem you hold; we will tell you whether the discipline fits it.
Inquire →Systems described on this page are research programs, not products. Praetu makes no performance claims for work in development, and nothing here is an offer for sale.