System Architecture

Rela AI is not a dashboard with disconnected modules — it's a pipeline where every piece feeds the next. This page shows the full system side-to-side so you understand how the parts you read about elsewhere in the docs connect.

Main pipeline

flowchart LR
  subgraph Ingest
    S1[MQTT sensors]
    S2[OPC UA / Modbus / S7 / EtherNet-IP]
    S3[HTTP webhook]
  end

  subgraph Predictive pipeline
    N[Field mapping<br/>+ normalisation]
    B[Baselines<br/>+ trends]
    D1[Anomaly detector<br/>IF + LOF]
    D2[Energy detector<br/>z-score + Page-Hinkley]
    D3[Prognostics<br/>RUL + CBM]
  end

  subgraph Consolidation
    A[Alert aggregator<br/>dedup by asset]
    Alerts[(_alerts)]
  end

  subgraph Orchestration
    H[AHI 5 sub-indices]
    M[Maturity 0..3]
    X[WhatsApp/Email agents]
    T[Tasks + SLA]
    CMMS[CMMS sync]
  end

  subgraph Delivery
    Report[PDF/Excel reports]
    Sched[Scheduled reports]
    CSAT[CSAT on chat close]
  end

  S1 --> N
  S2 --> N
  S3 --> N
  N --> B
  B --> D1
  B --> D2
  B --> D3
  D1 --> A
  D2 --> A
  D3 --> A
  A --> Alerts
  Alerts --> H
  Alerts --> X
  H --> M
  X --> T
  T --> CMMS
  X --> Report
  Report --> Sched
  X --> CSAT

Modules and where to read more

Design principles

Everything connects to everything

No isolated modules. A detection from the anomaly detector → goes to the aggregator → consolidates into _alerts → updates the AHI → can trigger an agent → that creates a task → that syncs with the CMMS. This is the differentiator over traditional IoT platforms that only show dashboards.

Zero hardware

The platform is 100% software. We don't sell sensors or PLCs — we connect to what the client already has via the 7 natively supported protocols.

Every tenant trains its own model

No generic predictive model. Each company trains its own with its own operational history — its baseline, its confirmed failures, its cycles. That's why the pipeline has maturity levels: it starts with basic alerts and evolves into confident prediction.

Shared canonical severity

The 3 detectors (anomaly, energy, prognostics) share the same info < warning < high < critical scale via adapters (severity.py). The aggregator uses this canonical scale to dedupe across systems. See glossary for the distinction with the legacy Machine Events scale.

config_version on every snapshot

Every document derived from the predictive config (_asset_health_snapshots, _asset_prognostics, etc.) carries the config_version active at computation time. This lets you answer "which thresholds applied when this asset was graded A?".

To get started

1. Read getting-started/concepts to understand maturity levels.
2. Follow getting-started/setup to connect your first asset.
3. Build your first agent with tutorials/first-agent.
4. Turn alarms live with tutorials/alarm-escalation.

See differentiation if you'd rather read why Rela AI differs from traditional CMMS + IoT platforms before you dive into tutorials.