Collatr

Possibly the most important article in our short series. How are SME manufacturers supposed to become more effective when the tools that might help them are beyond their reach?

In four days, Collatr Edge went from an empty repository to a working IIoT data collection agent. 10,600 lines of production code. 26,700 lines of tests. 191 commits. 11 build phases, each with independent code review. It reads Modbus TCP, OPC-UA, and MQTT. It processes and aggregates data at the edge. It stores locally in SQLite. It forwards via Sparkplug B. It serves a live dashboard from the device itself. It runs as a single binary on x86 or ARM. Every decision has a paper trail. Every phase has a review. Every bug found in review generated a rule that prevented the same class of bug in the next phase. Is it finished? No. There is a backlog. Modbus RTU over serial. EtherNet/IP. MTConnect. Fleet management. Predictive maintenance. The list is long. But the foundation is solid. The patterns are established. The code is tested beyond what most production systems achieve. This is what building in public looks like. Not a demo. Not a prototype. Working software, built from a real PRD, tested for real conditions, ready for real factories. More soon #collatr #digitalisation #servitization #AI #OT #closeTheLoop #buildInPublic

Collatr Edge has a built-in web UI. No cloud required. No external dependencies. It runs on the same device that collects the data. The stack: Datastar, Bun, server-sent events for real-time updates, ECharts for trend visualisation. All embedded in the compiled binary. No CDN calls. No npm install on the factory floor. The dashboard shows live metrics as they flow through the pipeline. Trend charts load historical data from the local store and append new readings in real time. CSV export includes both UTC and local timezone columns, because the person reading the spreadsheet thinks in local time. There is an OPC-UA certificate helper page. When the agent connects to a new OPC-UA server, it needs to trust the server's certificate. The web UI shows the certificate details and lets the operator approve it. No command line required. Seven spikes validated the stack before I wrote a line of production code. Datastar. Bun with JSX templates. SSE streaming. HTML fragment patching. Chart rendering. Static asset embedding. Full integration. The web UI is the product for most standalone and local network deployments. Built first-class, tested first-class. #collatr #digitalisation #servitization #AI #OT #closeTheLoop #buildInPublic

Most IIoT platforms assume a cloud connection. Collatr Edge assumes nothing. Three operating modes. Connected: full cloud link, real-time data forwarding, remote management. Local network: data stays on the factory LAN, accessible via the web UI, exportable as CSV. Standalone: the device talks to equipment and stores data. That is it. No network egress at all. The network policy is enforced at startup. Not at runtime, where a misconfiguration might silently leak data. The agent checks every output against the policy, reports conflicts, and refuses to start if the configuration violates the mode. Local network is the most common mode for UK SME manufacturers. Their data does not leave the building. They query it from a browser on the shop floor. They export it for their Monday morning meeting. That is the real use case. Not a degraded version of the cloud product. Standalone mode exists because some sites have no network at all. Just a device, a PLC, and a USB stick for data extraction. All three are in the code today. Tested. Reviewed. Working. #collatr #digitalisation #servitization #AI #OT #closeTheLoop #buildInPublic

Selling outcomes beats selling boxes. But most SMEs lack the data, telemetry and organisational readiness to move from products to service-based models...

Sparkplug B is a specification for MQTT in industrial environments. It defines how edge nodes announce themselves, how they report data, and how they handle disconnection. Collatr Edge implements the full edge node lifecycle. Birth certificates on connection. Death certificates (via MQTT Last Will) on disconnection. Periodic heartbeats with self-monitoring metrics. The Sparkplug B codec maps Collatr Edge metric types to Sparkplug types. Metric aliases for bandwidth reduction on constrained networks. Payload encoding with protocol buffers. The hub link session manager handles the state machine: offline, connecting, online, disconnecting. Graceful "going standalone" messages before cutting the connection. Automatic rebirth after broker reconnection. We extracted a shared MQTT client wrapper first. The MQTT consumer (input) and MQTT output now share connection logic. One set of reconnection behaviour. One set of tests. This is how CollatrEdge talks to the Collatr Hub. But it is also how it talks to any Sparkplug B-aware system. The protocol is open. The agent is not locked in. #collatr #digitalisation #servitization #AI #OT #closeTheLoop #buildInPublic

Collatr Edge ships as a single binary. No runtime to install. No dependencies to manage. `collatr-edge run` starts the agent. `collatr-edge config validate` checks your configuration before you deploy it. `collatr-edge config init` generates a template for your chosen mode: connected, local network, or standalone. The validate command checks per-plugin schemas. It reports network policy conflicts. It tells you, before the agent starts, that your standalone config references an MQTT output that the network policy will block. Structured JSON logging. Level filtering. Child loggers per plugin so you can trace exactly which input generated which error. A systemd unit file and deployment documentation. Because the person installing this on a Raspberry Pi in a factory is not the person who wrote it. Table stakes. Yet most IIoT tools still ship without them. #collatr #digitalisation #servitization #AI #OT #closeTheLoop #buildInPublic

Collatr Edge has 800+ tests. More lines of test code than production code. That is deliberate. The E2E suite runs a full pipeline: real config parsing, real plugin instantiation, real data flow from input to output. The sustained operation tests run thousands of metrics through the system and verify nothing leaks, nothing drifts, nothing silently drops. The buffer overflow tests fill the store-and-forward queue past capacity and verify the system degrades gracefully. Old data drops. New data keeps flowing. The agent stays up. The error resilience tests inject failures at every layer. A processor that throws. An output that refuses connections. A corrupt SQLite file. The agent recovers from all of them. I also tested a midnight UTC boundary bug that only triggers when timestamps span a day rotation. Found it in independent review. Would have been invisible in normal testing. Manufacturing software does not get to crash. The production line keeps running whether your code is ready or not. #collatr #digitalisation #servitization #AI #OT #closeTheLoop #buildInPublic

Raw sensor data is noisy. A temperature probe that reads 1,847 times in an hour does not need 1,847 rows in a database. It needs a min, max, mean, count, and standard deviation. One row. Collatr Edge runs processors and aggregators locally. Before the data leaves the device. The basicstats aggregator uses Welford's online algorithm for variance. One pass. Constant memory. No matter how many readings arrive in a window, memory usage stays flat. The filter processor drops readings that match configurable rules. Six evaluation types. Glob patterns for names and tags. The rename processor rewrites metric names and tags in flight. These run in a pipeline. Input, processor, processor, aggregator, output. Each stage isolated. If a processor throws an error, that metric is dropped. The rest of the pipeline keeps running. Edge processing is the difference between a useful system and one that drowns in data. #collatr #digitalisation #servitization #AI #OT #closeTheLoop #buildInPublic

UK manufacturing faces a skills crisis far deeper than a recruitment headache. 92% of manufacturing SMEs report skills gaps. The workforce is ageing toward retirement. Digital capabilities lag behind ambition. Most smaller manufacturers find themselves trapped: they need digital tools to improve productivity, but they lack the people to deploy or maintain them. The shortage concentrates in exactly the roles that matter most: maintenance engineers, process engineers, multi-skilled operators, and anyone who can bridge the gap between shop floor and software. Until the people problem is addressed, technology investment alone will deliver nothing.