Warning while there's still time to plan the work.
Detection at earliest observable onset
No failure labels, no model tuning, no new sensors
Air-gap capable, fully on-site
Recommend-only through fully autonomous
| A tool fails between scheduled PMs | Detects degradation early enough to plan the work |
| Calendar-based PM wastes good parts | Services on actual condition, not elapsed time |
| Alarm noise nobody can triage | Continuous health score, sensor attribution, urgency |
| Retiring expertise | Records every diagnosis as reusable institutional memory |
| Policy forbids sending data out | Runs entirely inside the facility |
Sense, judge, act — implemented as a five-stage cognitive loop:
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.
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.
| On-premises | Full system including AI reasoning on one GPU server inside your facility. Signed offline updates. Air-gap capable. |
| Hosted | Same platform, Imantics-hosted, for connected operations. Onboard in days. |
Failure caught ~4 days ahead. AUROC 0.851 · ~12% FPR
1,500+ high-risk parts flagged. AUROC 0.885
All modeled fault types detected — method verification
Regime change flagged days early — unsupervised detection
Load types separated, anomalous days isolated — zero labels
Eleven scenarios scored; offline model selected on evidence
Lead time scales with the failure mode: abrupt failures give days, gradual wear can give weeks.