When connected infrastructure fails, finding the fault is harder than fixing it.
Infrastructure should never simply be trusted on its own word. Iris is a hardware diagnostic that shows how a device is actually behaving. One source of evidence to find a fault, show a device meets its specification, or see where it is exposed.
A fault can sit in the device, the firmware, the network or the handoff between them, and no single team or vendor sees enough to place it.
The costly faults sit between the parts: each team checks its own, sees green and hands it on. No one can own what no one can pin down, and the person accountable is left to wait and chase, answerable for a fault they cannot reach.
EPRI, PNNL and TWAICE, BESS Failure Incident Database analysis, 2024
The factory floor tells the same story. Across more than 320 factory audits of grid-scale battery systems, nearly half of all quality findings were at the system level, in the integration and assembly of the parts, not in the cells or modules.
Clean Energy Associates, BESS Quality Risks Report, 2024 · 320+ inspections across 52+ factories, ~30 GWh
Each team is doing its job. The problem is structural.
Everyone sees a fragment. No one sees the system.
end to endNo one sees this
Facilities & operations
See equipment state and safety constraints.
Network teams
See connectivity, traffic and policy.
Vendors
See their own components, operating as expected.
Security & cyber
See threats, segmentation and policy. Not whether a control just broke a working device.
Leadership
Sees downtime, risk and pressure.
Each team owns its fragment. No one owns the whole, or the gaps where failures live.
During planned maintenance a pack is replaced and another added, and the work passes commissioning. A month on, the maintenance long forgotten, the first heatwave puts the site under load: packs begin shutting themselves down and the site derates, hard.
Operations sees packs dropping offline and cannot tell that a replaced controller, left misconfigured on the bus, is the cause, because confirming that means bus-level and configuration checks that sit with another team, and another set of tools.
Without a shared, cross-functional view of how the device is actually behaving end to end, investigation becomes sequential, slow and escalatory.
Output dips under heat
On the hottest part of the day a pack drops offline. The site derates for a moment, then steadies.
It spreads
Adjacent packs are running warmer and begin shutting down too. The derates cluster, with no obvious trigger.
The vendor loop begins
Each supplier checks their own box and finds it healthy. The packs report normal, and none of them explains the dips.
Workarounds replace understanding
The site is nursed through with derates and manual interventions. It runs, but confidence is low.
Root cause is never proven
A probable cause is agreed without proof. Reviews are uncomfortable, and teams move on knowing the same pattern will repeat under the next heatwave.
This is the gap Iris was built to close.
The cost is rarely just money.
These devices keep a ward monitored, a line running, a site powered and safe. When one fails and no one can place the cause, the harm lands first on whatever it was holding up. These estates are being asked to carry more of ordinary life every year. The evidence beneath them should be growing up just as fast.
Data centres are the corner of connected infrastructure where the cost of failure is actually counted.
Uptime Institute, Annual Outage Analysis 2026
Grid-scale storage keeps no such ledger yet. Its failures are counted; the money mostly is not, because a battery’s losses arrive as degradation, lost availability and disputes no one can settle.
And that is only the financial cost. While the cause goes unproven, the people who should be running the service are drawn into placing the fault instead. The reputation takes far longer to rebuild than the outage took to fix.
Iris reads the device itself.
Iris is a hardware diagnostic an engineer brings to the equipment, at commissioning, in maintenance or in the middle of a fault.
Connected through whatever interface the device exposes, it establishes how it behaves, how it is configured and what it talks to.
Not what the datasheet claims or the dashboard reports. What it is doing.
One connection runs the whole suite, against the device and everything around it: work that waits on four teams, done in a single visit. None of it is autonomous. Every step closer to the device is yours to take, never its own.
- 01
Learn
Iris watches the device in normal use and learns what healthy looks like for this unit. It acts on nothing.
- 02
Capture and compare
It records a device’s configuration, then on the next visit shows what has drifted since.
- 03
Listen
It reads the traffic between the device and everything around it, so a misconfigured neighbour has nowhere to hide.
- 04
Stand in
Iris emulates the device to see how each side actually responds rather than how it should.
- 05
ProbeDevice off the live estate
Iris tests the device for what an attacker could reach: stale firmware, weak hardening, default credentials, open ports. Almost no one can do this safely today, because the people who own security never get to handle the device. Iris lets the engineer who already maintains it do both, at planned maintenance.
What Iris may do is decided twice: once for the estate, in configuration; again at the device, by the engineer who knows what is live. Anything more active needs explicit authorisation, and every step is recorded.
Every test you switch off is certainty given up, and Iris says exactly what it can no longer establish. Switch off enough and you are back where you started: not knowing.
A device pushed hard off the live estate shows what it can safely take. A test you would not allow on day one becomes one you enable on evidence.
Which device or layer is at fault, whether it meets its specification, and whether it is exposed.
A device drifting from its documentation, or two units that should be identical behaving differently, before either becomes a failure. A fault found this early is the cheapest lesson an estate ever learns.
Engineer, network, security and vendor work from the same evidence, instead of four accounts that disagree.
The same examination that finds a fault shows the exposure: out-of-date firmware, weak hardening, default credentials, open ports.
The same evidence answers the Cyber Assessment Framework, confirms new kit will integrate before you buy it, and brings a fleet up the same way at commissioning.
Reliability grounded in evidence.
When you can’t get a straight answer.
Iris is built for connected, long-lived estates: many vendors, firmware that keeps changing, and a widening gap between what the documentation says and what the kit actually does. When something degrades, escalation is immediate and personal, landing on whoever’s nearest, regardless of where the fault actually sits.
Equipment kept in service for years, not replaced annually, assembled from many suppliers with mixed protocols, each running its own evolving firmware.
Behaviour that changes across firmware versions, configurations and conditions, while what’s written down stays stale, incomplete or simply wrong.
Intermittent, context-dependent failures that won’t reproduce on demand and fall between systems.
Kit that ships with weak logging, little security and no safe way to test it live, and falls outside the tools that manage everything else.
Maintenance windows are short and safety is non-negotiable. There’s no taking a live device down on a hunch, and no room to experiment on kit that’s still running.
We work as design partners.
We take on a small number of design partners at a time, working alongside your team, in your environment, on your terms, and honest at every step about what we know and what we are still proving.
- Bounded in scope and time.
- You shape it. How Iris works in your environment and the pace it earns are set with you, not handed to you.
We work a hard problem alongside the people who live it.
We have built systems where failure was not an option.
We have spent our careers building systems where a wrong reading has physical consequences.
Sensing and alerting for mines and nuclear facilities, control systems for refineries and logistics, and the software safety layer for a grid-connected battery system meant to keep the power on through a disaster.
With the hardware protections adapted out of that deployment, safety had to live in the software, which meant reading each battery by its own behaviour, not the summary it reported. That is how we found supposedly identical units were not identical at all, and it is the discipline Axions is built on: nothing is trusted because it reports itself healthy.
We have built and run systems in production at petabyte scale, across hardware, networks, data and machine learning.
Systems like that could not be bought off the shelf and bolted together; they had to be designed as one. It is why Axions can treat the device, the network and the data layer as a single engineering problem, rather than three teams pointing at each other. We read drift and decay as clearly as we read packets.
Three years of weekly fieldwork inside NHS hospitals, alongside the clinical engineering, IT, network and cyber teams.
The people who own a medical device know the device, not its networking or its security; the people who own the network and security know enterprise IT, rarely the IoT, let alone a device that can harm a patient. That skills gap is real and structural, and it is where the hardest faults sat, unowned. We worked across both worlds, often the ones who could get the teams talking and get to the bottom of a fault, and that is what Axions carries into other estates.
The problems we keep coming back to.
These are the problems we have chosen to work on: the ones we think matter most, get discussed least, and will only get harder as estates grow more complex and connected.
We’d rather listen than pitch.
30 minutes. Bring one fault you could never fully explain. We will talk through how Iris would have approached it, and tell you honestly whether we think it would have helped.
Start a conversation