Protect availability on the equipment you operate.

Warning while there's still time to plan the work.

What you get

Time to act

Detection at earliest observable onset

Nothing to configure

No failure labels, no model tuning, no new sensors

Data stays yours

Air-gap capable, fully on-site

You set autonomy

Recommend-only through fully autonomous

Problems it solves

A tool fails between scheduled PMsDetects degradation early enough to plan the work
Calendar-based PM wastes good partsServices on actual condition, not elapsed time
Alarm noise nobody can triageContinuous health score, sensor attribution, urgency
Retiring expertiseRecords every diagnosis as reusable institutional memory
Policy forbids sending data outRuns entirely inside the facility

How it works

Sense, judge, act — implemented as a five-stage cognitive loop:

Perceive
Remember
Reason
Act
Learn

Scores health from live sensor data · recalls similar past events · diagnoses the likely failure mode and urgency · recommends the action, or raises the work order in your maintenance system (SAP PM, Maximo) · records the outcome.

Autonomy is yours: recommend-only, approve-then-act, or fully autonomous. Every step is auditable.

ETCHER-02 · ion-beam sourceAlert
Failure mode
Ion-source filament approaching end of life
Root cause
Suppressor-voltage drift with rising beam-current variance
Action
Schedule source replacement within the next maintenance window (~4 days)
Confidence
0.82 · recommend-only
Reasoning trace recorded to the audit log
Product view — simulated data for illustration.

Connecting your machines

Consumes what your systems already collect — SECS/GEM and Interface A streams on semiconductor tools, OPC UA, historians, or file exports elsewhere. An evaluation starts from historical files alone.

Where it runs

On-premisesFull system including AI reasoning on one GPU server inside your facility. Signed offline updates. Air-gap capable.
HostedSame platform, Imantics-hosted, for connected operations. Onboard in days.

Evidence

Ion-beam toolreal data, blind

Failure caught ~4 days ahead. AUROC 0.851 · ~12% FPR

Parts-cleaning linereal production

1,500+ high-risk parts flagged. AUROC 0.885

Deposition chambersynthetic

All modeled fault types detected — method verification

Water networklive field data

Regime change flagged days early — unsupervised detection

Building energy250 days

Load types separated, anomalous days isolated — zero labels

Reasoning layerfour models

Eleven scenarios scored; offline model selected on evidence

Lead time scales with the failure mode: abrupt failures give days, gradual wear can give weeks.