Platform · Brief

From signal to decision,
in milliseconds.

Zoniax is a vertically engineered stack — sensor, edge, control plane, and SaaS — for plants that need to act on data, not just store it. The brief below is what we describe to engineering and IT stakeholders during a pilot kickoff.

01 · Approach

Two disciplines.
One operating picture.

We pair hardened physical instrumentation with online learning models so plant teams move from interpreting screens to acting on them.

i.

Sensors & Telemetry

Industrial-grade hardware engineered for the dust, vibration, and heat of real processing lines. Edge gateways stream readings to the platform at the cadence your operation demands.

  • Ruggedized sensors specified for IP65+ duty
  • Edge gateways supporting Modbus, OPC-UA, BLE, LoRaWAN
  • High-frequency timeseries with tiered hot/warm/cold storage
  • Configurable alert rules with hysteresis and dampening
  • Deployment models: cloud, on-prem, or air-gapped hybrid
ii.

AI & Intelligence

Online models continuously fit themselves to your plant's normal: flagging drift, recommending interventions, and learning from each operator decision they observe.

  • Anomaly detection trained per-asset, per-line, per-shift
  • Predictive maintenance ranked by expected downtime cost
  • Closed-loop quality scoring against batch specifications
  • Cross-sensor pattern recognition (process × environment)
  • Operator-in-the-loop labeling with full audit history
02 · Spans

Four spans, one contract.

Each span exposes a typed boundary the others depend on. Replacing a span — for compliance, for performance — should not invalidate the rest.

Span 01

Ingest

Schema-validated MQTT and HTTPS endpoints. Avro and Protobuf payloads. Replay buffering at the gateway, so nothing is silently dropped.

MQTT 5HTTPSAvroProtobufSchema Reg.
Span 02

Edge compute

Signed WASM modules execute deterministic rules at the edge. Drift detectors and quality scorers run beside the asset, not in the cloud.

WASMRustOPC-UALoRaWANBLE
Span 03

Identity

OIDC and SAML SSO. RBAC scoped to plant / line / asset / channel. SCIM for provisioning. Signed audit trails for every action.

OIDCSAMLSCIMRBACAudit
Span 04

Integration

Native adapters for the systems plants already run. Bi-directional where useful — work orders out, KPIs in.

ERPMESCMMSData lakeWebhooks
03 · Console

The plant, on one screen, not eight.

Operators see lines, assets, and signals together. Alerts arrive with context. Investigations are one click from raw timeseries.

A · KPI strip B · Event log · last 60 min
app.zoniax.com / plants / line-a / overview Sample data
OEE · 24h
87.4%
↑ 2.1 pp vs 7d
Yield · 24h
98.6%
↑ 0.4 pp
Energy · kWh/T
412
↓ 4.8% (good)
Open alerts
3
1 critical · 2 warn

Line A · Process temperature, 6h

12s sample

Event log · last 60 min

  • 14:42:08 Mixer 04: temperature rejoined band (−0.2 °C of target) LINE A · TEMP
  • 14:31:55 Pressure drift detected, within tolerance, monitoring LINE A · PRES
  • 14:18:02 CIP cycle completed, full reset of Pasteur. probes PASTEUR · CIP
  • 13:54:11 Bearing vibration spike, predicted MTBF revised −38h LINE B · VIB

Sample data. Your console mirrors your plant taxonomy.

04 · Deployment

Three deployments — same product.

Cloud

SaaS console on infrastructure we manage and operate. The fastest path to value.

Managed service Standard

On-premises

The same platform, deployed inside your network. Air-gapped variants available. Updates as signed bundles.

Containerized deployment Regulated industries

Hybrid

Sensitive telemetry stays on-prem, aggregated KPIs and ML training run in cloud. Selectable per channel.

Selectable per channel · per asset Most common
05 · Security

A short note on trust.

The platform is built around three default assumptions: the network is hostile, the operator is the source of truth, and every action eventually becomes evidence in an audit or incident review.

  • 01 Defense in depth. Encryption in transit and at rest, mutual authentication at the edge, and hardware-backed protection where the deployment supports it.
  • 02 Signed everything. Edge rule modules, firmware updates, and audit-log segments are cryptographically signed and verified before activation.
  • 03 No customer co-training. Models that operate on your plant are trained from your plant's data. We don't aggregate operational data into shared corpora without explicit research agreement.
  • 04 Right to leave. Documented exports for every customer-owned data class. Standard formats. No proprietary lock-in for the timeseries itself.
06 · Common questions

Frequently asked.

Do you require us to rip and replace existing sensors?

No. Most plants already have instrumentation that's serviceable; we instrument the gaps and bridge what's there. New sensors are only specified where coverage is missing or where existing hardware no longer meets the duty.

Does our data leave the plant?

Only if your deployment opts into the SaaS console. On-premises and air-gapped hybrid deployments are first-class: the same platform, with telemetry that never leaves your network and updates delivered as signed bundles.

How quickly does a pilot start producing signals?

Capturing live telemetry is the first milestone: a pilot is scoped to start producing data early, before any modelling. Anomaly detection follows once there is enough baseline data to learn from, and predictive-maintenance signals become reliable after an asset has been observed across a full maintenance cycle. We agree the specific milestones and timeline with you when the pilot is scoped.

What protocols are supported at the edge?

Modbus (RTU and TCP), OPC-UA, MQTT, BLE, LoRaWAN, and a small number of vendor-specific industrial buses. New protocol adapters are written in Rust and deployed as signed WASM modules.

Do you train models on our data?

Models that operate on a plant are trained from that plant's data, scoped to that organization. We do not aggregate customer data into shared training corpora without an explicit research agreement.

How much data do you need before a model is useful?

Enough to have seen the process behave normally, and enough to have seen it go wrong. That is why a pilot captures live telemetry before it models anything: anomaly detection follows once there is a baseline to learn from, and failure prediction becomes trustworthy only after an asset has been observed across a full maintenance cycle. We would rather scope those milestones with you than quote a number that sounds precise and isn't.

Do we need labelled data before you can start?

Not to begin. Anomaly detection learns a plant's normal without labels, which is usually the first model worth having. Labels make everything downstream sharper, so we build the labeling into the operator's workflow (operator-in-the-loop, with a full audit history) rather than asking you to hand over a labelled dataset up front.

Can model training run inside our network?

Where training runs is part of the deployment decision, not a default. On-premises and air-gapped deployments keep telemetry inside your network. The common middle ground is hybrid: sensitive telemetry stays on-prem while aggregated data and model training run in the cloud, selectable per channel and per asset.

What happens to a model when our process changes?

It goes stale, which is the normal condition of a model on a live plant, not a failure. A recipe change, a new feedstock, a replaced bearing: each shifts the baseline the model learned. Drift detection runs beside the asset rather than in the cloud, and retraining is then a decision with evidence behind it rather than a job on a schedule.

Bring us your
hardest line.

We'll spec a pilot scoped to a single line you already know is expensive — and we'll commit, in writing, to measurable signals in the first 30 days.