Private systems, in British rooms
Six systems built, one of them tested at my own desk every day. Each runs on hardware the client owns and leaves without needing me.
AI that stays in your building.
KREO: artificial intelligence engineering, data science and design architecture, Plymouth, UK.
The road to British excellence is built one room at a time.
One engineer, the whole stack. I design the system, serve the model on hardware you own, write the evaluation that proves it, and hand over the keys.
58pieces publishedsources on the page
6systems builtend to end
1in daily usetested at my own desk
0bytes to a third-party cloudby construction
Plymouth, UKUK time --:--:--Answered within one working day
01Selected Work2024 — 2026
AI desktop applicationKREO DesktopA local-first AI workspace: chat, browse, capture and research in one calm, dark instrument.Private index — client systemsSector, system and outcome only. Client systems run inside client walls, so discretion is part of the product; live viewings are arranged on request.
K-26-04Regulated sectorRetrieval · citation pipelineIn productionevery answer cited to its source document
K-26-02Marine & engineeringOn-prem inference clusterIn productionruns on client hardware; nothing leaves the building
K-25-11Esports & broadcastAgentic content pipelineShippedcontent drafted and routed by agents
K-25-07PropertyDocument intelligenceShippeddocuments read and answered in plain language
K-25-03EducationPrivate LLM serviceShippeda private model serving the institution
The work that came first
02The DisciplineSilicon to interface
Six layers, one engineer. Most people work one of them. The whole system holds together because the same person decides all six, and can defend each one with a measurement.
Drawn from the practice's stack: open-weight models on client hardware, retrieval over client documents, every answer traceable.
01Inference & servingOpen-weight models on hardware you own, quantised to the memory you actually have, measured in tokens per second and pounds per million.tok/s · £/M tok
02Retrieval & agentsAnswers that cite the document they came from. Agents that stop at the boundary, and ask before they act.citations · refusals
03Data scienceBaselines, confidence intervals and the honest error bar. Measurement before the claim, every time.p50 · p95 · n
04Systems & designComponent systems and tokens that stay usable as they grow. The interface is part of the engineering, not the decoration.tokens · states
05Motion & filmBroadcast, grade, sound and pacing. The craft that makes a system legible to a person in the room.frames · grade
06Writing & argumentFifty-eight published pieces, each carrying its sources. The reasoning is part of the build, not the marketing.58 pieces · sourced
03The MethodEvidence, or it did not happen
I do not trust a result I cannot re-run, a claim I cannot source, or a system I cannot hand over.
- Every number I publish can be traced to the source behind it.
- Every system arrives with the evidence that it does what was asked.
- Every handover is complete enough that your team does not need me.
- Every claim that turns out wrong is corrected in public.
$16.8bnAnthropic, third quarter 2026Confirmed against two independent reports. Labelled as reported, not audited.
04The RoadWhy this, and why here
Britain does not have a compute problem. It has a tenancy problem.
The models are rented. The compute is leased. The work leaves the building and comes back as a subscription. Two years of writing about what that costs, and the answer does not change: the capability goes, and the invoice stays.
The correction will not arrive as a national project. It arrives room by room, from people who can hold the whole stack and will not hand the keys to someone else's cloud.
My part is narrow. Build systems that stay here. Publish the reasoning so it can be checked. Keep the standard high enough that “British” means something in the room.
A standard that can be checked
Every claim carries its source; every build carries its evaluation. The record stays public so the standard can be argued with, not just asserted.
The room, not the monument
A practice that trains the next pair of hands and leaves capability behind it. British excellence is a habit of work, not a press release.
The journal is the argument. Every figure goes to a source, every source is linked, and the pieces I got wrong stay on the page with the correction attached.
23 OctLeaders, and Nations16 min read19 OctOne Bad Host4 min read19 OctThe Front Page Is a Promise4 min read
Before the AI years
Five stars earned in design and film, with clients across esports, broadcast and games. The AI work stays quiet because it runs inside client walls. Ask for a reference call and, where the client permits, I will arrange it.
“Awesome guy to work with and insanely talented. Recommend 200%”
06QuestionsPlain answers
What does KREO actually build?
Private AI systems: open-weight models served on hardware you own, retrieval over your own documents with every answer cited, agents that stop at the boundary, and the interfaces that make the whole thing usable. One engineer works across every layer, which is why the parts fit.
What kind of firm is this for?
Firms that cannot send their data to someone else's cloud: regulated industries, marine and engineering, media and broadcast, and any team that needs the system to run inside its own walls. If a hosted API is genuinely the better answer, I will say so in writing rather than build something you do not need.
Where is KREO based?
In Plymouth, Devon. I live and work in the Greater Plymouth area, and I was born and raised in Greater Manchester, where I keep strong working ties. Engagements run on-site across the South West, in Manchester, and remotely throughout the United Kingdom.
Can AI run privately on our own hardware?
Yes, and that is the whole point of the practice. Open-weight models on your GPUs or private cloud mean no data leaves your environment and no dependency on a third-party API. For regulated or client-confidential work it is usually the deciding factor.
Who does the work?
Brandon Allen, directly. KREO is an independent practice, not an agency: one accountable engineer across architecture, data science, implementation and deployment, with every system documented and handed over so your team can run it without me.
Can I see examples of the AI work?
Not publicly. The systems are built for client environments and run on client hardware, so they are not exhibited here. Live viewings are arranged on request: tell me your sector and I will show you systems close to it, and talk through how the same engineering would apply to yours.
How does working together begin?
With one email. I reply with what I would build, what it would cost, and whether it is worth doing. If it is not worth doing, I will say so, and you will have that in writing. When it is, the first paid step is a one-week audit: a written assessment you can act on, or take elsewhere.
07The DoorDirect line
If you are building something that has to stay yours, write to me.
I will tell you what I would build, what it would cost, and whether it is worth doing. If it is not worth doing, I will say so, and you will have it in writing.
BrandonAllenAI Engineer · Data Scientist · Design Architect
BA26Struck at Plymouth · to KREO standardPlymouth · England
Or write it here
Or pick a time in the diary
KREO booking desk
Book a studio call
Choose a focused slot and hit book. Your request is emailed straight to me at brandon@kreostudio.co.uk, so it lands in my inbox rather than disappearing into a third-party widget.
UK time. Weekday slots include a built-in handover buffer.
