The market is moving

Fabs are changing — from Automatic to Autonomous.

Chipmakers have committed to fabs that run unattended by 2030 — and every tool inside them will have to understand its own health.

Predictive maintenance your equipment carries into every customer site — no labeled failure histories, manually tuned thresholds, or per-failure model engineering, and no data leaving the facility.

Your customers are specifying autonomous fabs. Your tools will be judged on whether they can support one.

The industry's stated direction is fabs that run unattended by 2030. Early unmanned production lines are already reporting equipment failures cut by roughly 90% and overall equipment effectiveness more than doubled. Fabs are converting that experience into procurement requirements: a tool that cannot report and reason about its own health becomes a tool that needs a person standing next to it.

Agent Athena™ is how your equipment carries that capability — without you building an AI organization to do it.

What it delivers

Lower service cost

Emergency dispatches become scheduled visits. Unplanned warranty events become planned work.

A differentiated product

Autonomous predictive maintenance becomes a capability of your equipment — offered to your customers in the form you choose.

Fleet visibility

With customer consent, reliability insight from the field without breaching the data wall.

Why it scales where other PdM hasn't

Conventional PdM Agent Athena™
Labeled failure history per modeLearns each tool's own normal
Engineering hours at every siteNo failure-model training or threshold tuning
Add-on sensors to installUses your existing sensors
Cloud connectivity, offsite analystsRuns on-site, air-gapped

The end-to-end workflow

A maintenance event that runs itself: detected, diagnosed, ordered, staged, scheduled — one workflow, on one platform, inside the facility.

Detect
Running today

Health scored continuously from the tool's own sensors; anomaly raised at the earliest physically observable onset

Diagnose
Running today

Failure mode, urgency, recommended action, with per-sensor attribution and a full reasoning trace

Stage
Roadmap

Parts reserved and kitted against the predicted window

Schedule
Roadmap

Tool time booked into the maintenance window that fits

Every deployment begins with detect and diagnose in recommend-only mode. The execution stages follow as you raise autonomy — per tool, per action, at the pace your processes allow. Every step is written to an audit log.

Warning arrives at the earliest physically observable onset — days to weeks depending on the failure mode.

Health index 80 onset breach Attribution Sensor 1 Sensor 2 Sensor 3 Sensor 4
Product view — simulated data for illustration.

Data in, work out

InConsumes the SECS/GEM and Interface A (EDA) streams your collection systems already gather; OPC UA, historians, or file export elsewhere. A POC needs only historical files.
OutRecommendations to your service workflow. At higher autonomy, work orders into the customer's maintenance system (SAP PM, Maximo) — duplicate-safe.

Built for the fab floor

Your tools run air-gapped. So does Agent Athena™.

Fully local AISensing and reasoning on one on-site GPU server. No cloud calls.
Offline updatesSigned bundles. No phone-home, no license server.
Data stays putModels train on-site, per tool.
Isolated tenantsRole-based access, full audit trail.

Air-gap isn't a limitation we tolerate. It's the deployment we engineered for.

Evidence

Ion-beam process toolreal production data, blind

Trained on early healthy operation only, then scored a month of history blind against a known ion-source failure. Caught it ~4 days ahead, at the earliest onset visible in the tool's real signals.

AUROC 0.851 · ~12% FPR

Precision parts-cleaning linereal production data

Surface-quality signatures across a high-volume cleaning operation, scored without failure labels. Identified 1,500+ high-risk parts for review before they reached the next process step.

AUROC 0.885

Deposition chambersynthetic simulation

A controlled benchmark with modeled fault types injected, used to verify the method against known answers. All modeled fault types detected.

Method verification, not a production result

Water infrastructure networklive field data

Trained on the network's early normal operation, then run forward blind. Flagged a coherent regime change days before it was operationally obvious — no labels available, so reported as detection only.

Unsupervised detection

Building energy systems250 days of real data

Power-quality and consumption data with no labels of any kind. The method separated load types on its own, characterized seasonal behavior, and isolated anomalous days.

Zero-label categorization and anomaly detection

Reasoning layerfour models benchmarked

Eleven engineering scenarios scored across four local language models for structured-output correctness and action quality. A permissively-licensed model was selected on the evidence and runs fully offline.

Selected model: 0.836 overall

Evidence at a glance

Ion-beam process toolReal production, one month
Protocol
Blind run-to-failure; trained on early healthy data only
Result
~4-day lead · AUROC 0.851 · ~12% FPR
What it establishes
Detection and lead time on a real tool
Parts-cleaning lineReal production
Protocol
Unlabeled scoring, verified against outcomes
Result
AUROC 0.885 · 1,500+ parts flagged
What it establishes
Detection on production quality signatures
Deposition chamberSynthetic simulation
Protocol
Injected modeled faults
Result
All modeled fault types detected
What it establishes
Method behavior — not customer validation
Water network (HydroView)Live field data
Protocol
Trained on early normal, run forward
Result
Regime change flagged days early
What it establishes
Detection only — no labels to score diagnosis
Building energy250 days, real
Protocol
Fully unsupervised
Result
Load categories separated, anomalous days isolated
What it establishes
Zero-label generalization across domains
Reasoning layer11 engineering scenarios
Protocol
Four local models benchmarked
Result
Top model 0.836, runs offline
What it establishes
Model selection evidence — not field accuracy

We label every result by what it does and does not establish.

Offered your way

Embedded or white-label · bundled through your channel · a service tier for your install base. Your decision, shaped together.