Astyr — AI fault diagnosis for factory equipment

Diagnose factory breakdowns like your best maintenance engineer.

Astyr captures the fault-diagnosis reflexes of a plant's most experienced engineer and runs them on every breakdown — grounded in the factory's own documents, and verified by the operator on the floor before anyone touches the machine.

Built forManufacturing·Energy·Utilities·Data centers

When a line goes down, the clock starts — and the diagnosis is the slow part.

Detecting a stoppage is easy. Diagnosing the root cause and prescribing a safe fix is where the minutes bleed — a technician hunting the cause across manuals, schematics and years of past breakdowns, often on a night shift, near energized or moving equipment. The person who reads the fault instantly is a veteran, and there are fewer of them every year.

$1.5Tper year

Unplanned downtime cost to the Fortune Global 500 — about 11% of revenue.

Siemens / Senseye, True Cost of Downtime 2022

>$2Mper hour

What a single hour of downtime costs in automotive manufacturing.

Siemens / Senseye, True Cost of Downtime 2022

2.1Mby 2030

U.S. manufacturing jobs that could go unfilled — a ~$1T cost in 2030 alone.

Deloitte & The Manufacturing Institute, 2021

~44median age

The U.S. manufacturing workforce. When a veteran retires, the diagnostic instinct leaves too.

U.S. Bureau of Labor Statistics, CPS

A team of agents, with the operator in the loop.

Astyr is not one model answering once. A correlation agent localizes the fault; a root-cause agent ranks probable causes with a cited confidence score — and below its threshold it refuses to guess and re-retrieves. Then the operator takes a real field measurement. If it fits, the cause is validated; if it contradicts, the agent pivots and re-diagnoses live. Only then does Astyr issue the cited, step-by-step procedure — safety and lockout steps inline.

The physical validation loop is the safety guarantee. Astyr never issues a repair action on an unverified diagnosis. It hands the operator a hypothesis and a safe test — not an instruction to act.

confidence-gatedrefuse to guessoperator-verifiedcited procedure

Not “agents work.” A narrower, more defensible window.

The generic why-now was the starting gun the whole agentic wave heard at once. Ours is specific — and we can demonstrate we clear it.

A specific agent bar just became clearable.

Not “an LLM answers.” Plan, retrieve behind a confidence gate, refuse to guess when unsure, request the exact missing document, call tools — and re-diagnose live when a physical measurement contradicts the hypothesis. Retrieve-then-answer copilots don't clear that bar.

The knowledge-capture window is closing.

Encoding a veteran's reflexes is only possible while that veteran is still on the floor. The retirement curve and a ~44-year-old median workforce mean that window is open now and narrowing.

We can prove it on real industrial data.

Manufacturing is the wedge because it is where we can start building the validated-incident corpus on real, messy factory documents — not curated demo data.

Diagnose factory breakdowns like your best maintenance engineer.

When a line goes down, the clock starts — and the diagnosis is the slow part.

A team of agents, with the operator in the loop.

Not “agents work.” A narrower, more defensible window.

A specific agent bar just became clearable.

The knowledge-capture window is closing.

We can prove it on real industrial data.

Judged by the people who build the models.

Anthropic “Built with Opus” 2026

RAISE Summit 2026 · Vultr track

See it run on a live incident.