Why Rela AI?

Rela AI is not yet another CMMS, nor a generic IoT platform, nor a digitised HACCP app. It is a Plant OS for SME food makers — a conversational operating system for industrial bakeries, dairies and small-to-mid-sized meat plants.

This page documents the four differentiators that make the category defensible. Each one answers a concrete question: "what does Rela do that competitors cannot, and why can they not simply copy it?".

Rela AI is not a dashboard waiting for someone to look at the data. It is a plant agent that listens to your PLC, runs HACCP in real time, sustains the cold chain and orchestrates your team over WhatsApp — all on top of a single asset and process graph.

The four differentiators:

1. Reading existing PLCs over VPN — no hardware of our own
2. Zero-history bootstrap — extruder from the first month
3. Conversational agent layer — WhatsApp and email as primary UI
4. Shared asset / process graph — seven modules, one reality

---

1. Reading existing PLCs over VPN — no hardware of our own

The problem

The typical SME food plant already has controllers in place: Siemens S7, Schneider Modicon, Rockwell ControlLogix, Carel, WAGO, OEM controllers from cold-room and tunnel makers. Each one exposes data over Modbus TCP, OPC UA, S7, EtherNet/IP, MQTT or proprietary protocols. The operator already has the manufacturer's HMI. The problem is not lack of data: those data live in silos and nobody reads them outside the HMI.

How Rela solves it

Rela connects an industrial VPN to the plant network and subscribes its listeners to the existing PLCs — Modbus TCP polling, native OPC UA subscription, MQTT through a broker (including Sparkplug B and OPC UA Pub/Sub), S7 and EtherNet/IP half-duplex polling. No hardware of our own: no sensors to buy, no edge gateway to maintain, no extra maintenance burden on your team. A single VPN serves multiple controller brands in the same plant.

Augury and Tractian are sensor-first: they need to sell and install their accelerometers to start. That excludes them from process variables (oven temperature, pasteuriser pressure, proofer humidity) that already live in the PLC.

Proof

• COLIP case — Carel proofer fleet over Modbus: 35 proofers from four different brands, a single VPN, native Modbus reads with no extra sensors.
• Multi-brand Modbus on a single VPN: Schneider + Carel + WAGO in the same bakery plant, without touching the original HMI.
• Multi-machine on a single tunnel: the architecture of the VPN + listener manager.

Competitors that fail here

• Augury and Tractian: require their own sensor; they do not read the original PLC.
• GE Predix, Siemens MindSphere: require their edge gateway or specific OPC UA stack; they do not scale down to SMEs.
• Classic CMMS (SAP PM, Maximo, Fracttal): do not read PLCs at all, rely on manual capture.

---

2. Zero-history bootstrap — extruder from the first month

The problem

The industrial predictive-maintenance standard says you need three to six months of clean history to train a model. That disqualifies every brand-new plant, every freshly installed asset, every machine whose previous owner never exported data. SME food makers rarely have clean history: what they have is three years of badly labelled stops on a spreadsheet.

How Rela solves it

Rela starts in a bootstrap mode that combines:

• Physical rules from the asset manual (RUL ceiling per temperature, vibration limit, nominal current).
• Pure change detection (Page-Hinkley on energy and vibration) that needs no learned baseline.
• Z-score with partial baseline that calibrates shift by shift.
• Canonical severity (info < warning < high < critical) — for the first week only warning and above are published; the operator qualifies each one and that feeds back into the baseline.

By day 30 a brand-new asset already has reliable AHI and RUL. By day 90 the model is indistinguishable from one trained on two years of history.

Proof

• Zero-history bootstrap — extruder: brand-new extruder with zero data, first useful alert at day 11.
• RUL ceilings and rule hierarchy configuration: technical documentation of bootstrap mode.
• Anomaly detection — ML ensemble: how the ensemble combines Page-Hinkley + z-score without a complete baseline.

Competitors that fail here

• Augury: documented cold start of 90–180 days per asset.
• Tractian: same — sensor-first and baseline required.
• CMMS: does not do predictive maintenance, only scheduled preventive.

---

3. Conversational agent layer — WhatsApp and email as primary UI

The problem

The shop-floor operator at an SME food plant does not have a laptop on the line. They have a phone, dirty hands and 18 seconds to decide what to do when oven HR-3 spits smoke. SafetyChain, Tulip, Augury and the classic CMMS all assume that the primary UI is a dashboard. For the operator, a dashboard is a place they almost never visit.

How Rela solves it

Rela's primary UI is WhatsApp. The operator talks to a conversational agent that understands natural language, accesses real-time PLC state, recommends actions and opens work orders for the team to approve. The same agent handles email, PDF reports, calendars and Postmark when the context calls for it.

The agent's tools cover four categories:

• Data query — find an asset, a lot, a technician, a CCP, a work order.
• Internal actions — generate PDF reports, pre-assign tasks, draft messages, locate personnel.
• External connections — call REST APIs, subscribe to MQTT topics, read OPC UA, sync with CMMS or ERP.
• Human approval — every action that touches a piece of equipment or a team member needs approval before running.

Proof

• WhatsApp agent — create and Create an email agent.
• Unified alert inbox: how the agent consolidates alerts from multiple detectors into a single conversation thread.
• Unified alert inbox case: an end-to-end with three detectors firing, a single message in WhatsApp, a single technician assigned.

Competitors that fail here

• SafetyChain: dashboard-first, does not solve the problem of the on-line operator.
• Tulip: configurable mobile apps, but they require the operator to open the app — no conversational agent, no WhatsApp.
• Augury / Tractian: notify via their app or email, with no agent that can answer a question from the operator.
• Classic CMMS: zero conversational layer.

---

4. Shared asset / process graph — seven modules, one reality

The problem

The SME food maker that tries to digitise ends up with five separate products: a CMMS for maintenance, a SafetyChain app for HACCP, the cold-room manufacturer's app for cold chain, an Excel sheet for OEE, a paper folder for LOTO. Each product has its own version of the asset "oven HR-3". Reconciling them eats hours every week and still never returns the same answer.

How Rela solves it

Rela's seven layers — conversational agent, HACCP, cold chain, OEE, maintenance, LOTO, data and integrations — share a single asset / process graph. Oven HR-3 is a single node in the graph, with edges to its HACCP CCP, to its downstream cold chain, to its OEE line, to its maintenance work orders and to its LOTO points.

This is not just "good data design": it is the only way for the WhatsApp agent, on receiving "oven HR-3 stopped", to know simultaneously which CCP is at risk, what OEE loss line 2 is accumulating, what preventive work order was scheduled and which LOTO must be executed before opening the chamber.

Proof

• Plant OS — what it is and why your plant needs it: the asset-graph diagram with the 7 layers and the full argument.
• Architecture: shared data model and how it is exposed to agents.
• Unified alert inbox case: a concrete example of how a cross-cutting alert (HACCP + maintenance) gets resolved through the same channel.

Competitors that fail here

• CMMS + SafetyChain + cold-chain app + OEE Excel: four tools, four asset models, zero unified answer.
• Tulip: single platform, but generic — no deep HACCP and no food-specific asset graph.
• Augury / Tractian: maintenance only; no HACCP and no cold chain in their graph.

---

Summary

The combination is what matters. Any competitor could copy one of the four; copying all four at once requires being vertical food + PLC-reader + agentic + compound from day one.