The Lab
The Department of Tomorrow.
Working prototypes, not promises. What we have built, what broke, and what it taught us about delivering AI.
The Lab
Working prototypes, not promises. What we have built, what broke, and what it taught us about delivering AI.
Why this page exists
The quickest way to learn where a technology breaks is to break it yourself. Everything filed here is a working prototype, built by us with AI coding tools in days rather than months, and used for real.
None of it is enterprise grade, and each file says why. That gap, between a prototype that works and a service you can stake a business on, is where most of our advice lives.
Psst. Press ` anywhere on this site, or open the terminal.
An AI receptionist that answers our phone line, takes messages and hands callers to a person with a short briefing, so nobody has to repeat themselves.
Callers speak to Illiriya in a natural British voice. It answers questions about the firm, takes messages, recognises returning callers and passes them to a person with a summary of what they need. Texts work too.
Before choosing a model we ran eight of them through the same scripted calls. Several promised to take a message and then never did; one read its own reasoning aloud to the caller. The model that got the job was the one that did what it said, quickly.
In testing, an early model told a returning caller what their previous call had been about. It was a test, but it was exactly the failure a compliance team loses sleep over.
We tightened the rule to "never repeat what earlier contact was about, even if asked", tested every model against it, and made holding that line a condition of the job.
Writing a privacy rule into a prompt is not the same as enforcing it. Test the rule against the model, adversarially, before a customer does it for you. And treat model choice as a governance decision, not a procurement one.
This is one phone line, built in days. A regulated contact centre needs resilience, recording policy, workforce management and audit that a prototype does not have, which is why platform choice still matters.
Text replies reach UK mobiles only. Calls and messages are handled as set out in our privacy policy.
Seven releases in under three weeks, from first calls in the browser on 21 September to voicemail, smart routing and quality scoring on 9 October. Taken from the app's own release notes.
Nobody hits a dead end any more, calls reach the right person first, and every conversation gets a score.
A second pair of hands on every call, and one page for everything you know about a caller.
A new name and look, and the app now knows who is really at their desk.
Illiriya learns about the business, and the team can see the whole floor at a glance.
Hold, transfer, callbacks and a phone in your pocket.
An AI receptionist who answers every call and text.
Day one: a contact centre in the browser.
Our own operating software for a pocket computer, turning it into a terminal for Claude and our own AI agents that keeps working over a radio mesh when there is no WiFi.
Illiriya OS runs on an M5Stack Cardputer, a keyboard and screen smaller than a phone. It pairs with Claude, talks to two AI agents running on our own machines, and records voice notes that are transcribed on our hardware rather than in the cloud.
It also speaks Meshtastic, an open long-range radio mesh. When WiFi drops, a question to an agent travels by radio to a gateway, and the answer comes back the same way.
Version 0.5.0 restarted itself three seconds after every boot. The cause was a time-zone function, called on every cycle, that slowly ate memory on a device with very little to spare.
We wrote our own time-zone handling, checked it against more than 368,000 cases, and added a safe mode and a crash log so that the next failure would explain itself.
Agents do not have to live in a browser tab, or even on the internet. Ask where the work actually happens, on a shop floor, in a vehicle or out in the field, and design for the connection you will really have.
And build the failure path first. A safe mode and a log are what turn a crash into a fix rather than a mystery.
This one lives in a pocket, so there is no public demo. Photographs are on their way.
Twenty-seven releases in nine days, from first boot on 30 September to working away from WiFi on 8 October, grouped here by version. Taken from the device's own change log, with a few security and radio tools left out on purpose.
The agents keep working when the WiFi does not.
The device reaches our agents with no WiFi at all, and learns to look after its own memory.
Voice notes that never leave our network, and a device that doubles as a clicker.
Private conversations over the mesh, checked against the published standards.
Instruments for the radio mesh, and a proper GPS.
Built, installed and running on the device on the same day.
Five AI models from five labs debate a question over three rounds. Every claim must cite its evidence, and a different model checks it.
The Council draws on a year of AI research, laid out as an evidence galaxy you can explore. For each question, the 45 most relevant items become the evidence pack, and the models may argue from that pack alone.
After each round, every claim goes to another model, which reads the cited sources and rules it supported, partly supported or not supported. Anything unsupported is challenged and has to be answered.
In one debate the entire rebuttal round failed: none of the five models returned a usable answer, and one checker returned no verdicts at all.
We published that transcript with the gaps showing, then made the engine sturdier and sent the question back for a full rerun. A system that hides its own failures is not one you should trust with your evidence.
A single model's answer is an opinion. Several independent models, held to the same evidence and checking each other, get much closer to something you can defend to a board or a regulator.
The design choices that matter are the unglamorous ones: a fixed evidence base, rotation so that nobody marks their own homework, and a visible record of what failed.
Built in two days: the galaxy and the debating engine on 7 October, then a wider library, a star chart and a model league on 8 October. Taken from the project's own history of 23 changes.
More sources, a flat map of the field, and every region named.
Which model's claims hold up, and a record of every failure.
A debate every Saturday, and an engine that no longer loses an hour to a bad night.
Watch the argument happen among the evidence.
Five models, cited claims and a different model checking each one.
A year of AI research, laid out as a map you can fly through.
We add to the Department whenever something we have built teaches us something worth passing on. If you would like to know how any of it would work in your organisation, we would be glad to talk.