The fire that doesn’t happen
is worth more than the data about it.
Every mains-powered Justworx device sits on the customer’s electrical system. That position turns out to be worth more than anything it was installed to do.
Faults announce themselves.
Almost nothing is listening.
Electrical arcing faults precede a significant share of structural fires, and they’re detectable well before ignition — a characteristic signature in the waveform, sometimes hours ahead.
Detecting them requires something watching continuously, sub-cycle, at the point of use. That is not what a smart meter does, and the reason is bandwidth, not capability.
A modern meter samples fast — some monitor power quality at over 14 kHz and track harmonics to the 63rd order. Then it transmits roughly 3 kilobytes a day. Full waveform capture would be more than a gigabyte per meter per hour, which as one major manufacturer put it is “clearly too much data to return over an RF mesh network.” Voltage typically reports at fifteen-minute intervals sent every four hours. Outage notification is optional, delayed, and vendors deliberately throttle it during storms — it degrades precisely when it matters most.
The data exists at the meter and dies there, because no pipe is wide enough to carry it.
An edge device doesn’t need the pipe.
Put the intelligence where the measurement happens and the bandwidth problem disappears. The device watches the waveform continuously, recognises the signature locally, and sends the conclusion.
A few bytes that say this circuit is about to fail — instead of gigabytes that say here is every waveform, you work it out.
That’s the entire architecture, and it’s the same pattern every Justworx product already uses: sense locally, decide locally, transmit only what matters.
The buyer isn’t the utility.
It’s whoever carries the fire risk.
This distinction is the difference between a business and a science project. Utilities buy slowly, procure through multi-year cycles, and have shown limited appetite for third-party grid data. Insurers buy things that reduce claims, and they buy them immediately when the arithmetic works.
The arithmetic here is unusually clean:
Insurers distribute it because it saves them more than it costs.
No token. No yield. No ROI story. An insurer buys a device because it prevents a claim.
The hardware goes into the building anyway.
The expensive part of that model is getting hardware into buildings. It’s also the part that happens on its own when partners ship Justworx products, for entirely unrelated reasons.
A SmartSwitch installed to control a geyser is, incidentally, a grid-edge sensor in exactly the right physical position. Sentinel doesn’t require a deployment programme. It rides on the devices going into those buildings anyway.
And the wider context is stark:
Distributed fault networks have detected faults hours before major fire ignitions. The problem is enormous and the instrumentation is close to absent.
The section that makes the rest believable.
- It needs hardware we haven’t shipped yet. Sub-cycle waveform sensing requires an analogue front end and sampling rates beyond what our current modules do. Series 3’s SPI and I²C make an external front end tractable, but this is a hardware development programme, not a firmware update. We’re being explicit about that because the alternative is implying a capability we don’t have.
- Classifying arcs on-device is genuinely difficult. Distinguishing a dangerous series arc from a light switch, a vacuum cleaner or a motor start is exactly the problem that makes arc-fault detection hard. False positives destroy trust with insurers faster than missed detections do. This is a real machine-learning problem, on constrained hardware, and we don’t intend to understate it.
- Insurance is a slow channel with a high evidentiary bar. Carriers want peer-reviewed loss data before they underwrite anything. That means a field programme and a claims study before revenue — a multi-year path, not a product launch.
- We’re not selling grid data to utilities. The comparable network has marketed grid insight to utilities for years without a single named utility customer. We’d rather learn that secondhand than repeat it.
Honest about the distance.
Mains-powered devices that sit in exactly the right physical position, on-device inference, and a network that carries small payloads efficiently — with the fleet still early and growing.
An analogue front-end reference design and an arc-signature model trained on real fault data.
A field programme with an insurance partner, and the claims evidence that turns a capability into a channel.