A small, senior AI studio in London
BrashPixels builds production AI systems for founders and product teams — the kind that survive contact with real users, real data and a real invoice at the end of the month.
In one paragraph
BrashPixels is a London AI development studio founded by Joe Scarffe. We design and ship production LLM applications — agents, retrieval systems and AI product surfaces — on Anthropic Claude, OpenAI GPT and open-source models, plus automation that plugs into the tools teams already use. Everything we build is delivered into the client's own repositories and owned by the client. We price each engagement against its actual scope — how that works.
Why the studio exists
There is no shortage of people who can get a language model to do something impressive once. The gap in the market is between that and a system a business can depend on — one with an evaluation harness, a cost ceiling, an audit trail, and a handover document that means the client is not permanently dependent on the people who built it.
Most AI projects do not fail on model capability. They fail because nobody defined what "working" meant as a number, because the data turned out to need three weeks of archaeology nobody budgeted for, or because the running cost surprised someone in finance four months in. Those are engineering and scoping problems, and they are what we organise the studio around.
Who you actually work with
BrashPixels is deliberately small and senior. The person who scopes your engagement is the person who builds it — there is no layer of account management between the conversation and the code, and no junior team inheriting a specification written by someone who has moved on.
That structure has an obvious cost, which we would rather state than hide: we take a limited number of concurrent engagements and sometimes cannot start when you need us to. It also has the benefit you are actually buying, which is that recommendations are made by people constrained by having to implement them. It is much harder to propose something unbuildable when you are the one who will be building it.
How we work
Evidence before build
Every engagement starts by establishing what "working" looks like as a measurable number. If nobody can define it, that is the first thing we fix.
Vendor-neutral
Claude, GPT, Gemini and open models are all on the table. We design so the model is swappable, because the field moves too fast to marry one.
You own everything
Code, prompts, eval suites and documentation land in your repositories. No licence, no proprietary layer, no lock-in by design.
Costs forecast up front
Model, hosting and infrastructure spend is projected during discovery — paid by you directly, never marked up by us.
What we won't do
A short list, because it is more informative than a longer list of things we will do.
- Build an agent when you need a workflow. Non-determinism is a cost. We recommend the simpler architecture when the simpler architecture wins, even though it is the smaller invoice — the reasoning is set out on our agent development page.
- Quote before seeing your data. Data readiness is the single largest cost driver in AI work. A number produced before anyone has looked at it is a guess.
- Mark up your token spend. A supplier who profits from API usage has an incentive to make your system chattier. You pay providers directly.
- Ship without evaluation. If there is no harness, there is no way to know whether the next change helped, and no way to catch quality decay before your users do.
- Sell consultancy that only leads to more consultancy. The roadmap should be executable by your team without us.
Who we work with
Founders and product teams, mostly between seed and Series B, plus established SMEs with a specific operational problem worth solving properly. Sectors we work in regularly include fintech, professional services, e-commerce and media — and we have a working familiarity with the extra diligence that regulated environments require.
Geographically: London and the UK in person, Europe and the US remotely. Most engagements run remote-first with in-person workshops at the points where being in a room genuinely helps, which is usually discovery and handover.
What we build
Four practices, one team:
- AI development — production LLM applications, structured outputs, evals and AI product engineering.
- AI agent development — tool-using agents with real audit trails and iteration limits.
- RAG development — retrieval over your own documents, with citations and permission-aware search.
- AI consultancy — opportunity mapping, technical due diligence and roadmaps you own.
- AI automation — copilots and workflow automation inside the tools your team already uses.
We also write about the work in Insights, mostly the parts that are inconvenient to admit — what things cost, what does not work, and where the received wisdom is wrong.
Frequently asked questions
Who is behind BrashPixels?
BrashPixels was founded by Joe Scarffe and operates as a small, senior studio in London. Engagements are staffed by the people who scoped them — the person who writes your roadmap is the person who builds it, which is the main structural difference between a studio and a consultancy.
How big is BrashPixels?
Deliberately small. We run a senior-only team and take a limited number of concurrent engagements rather than scaling headcount, because AI delivery quality depends far more on judgment than on hours available. It means we occasionally have to say no to timing.
Where is BrashPixels based?
London, United Kingdom. We work with clients across London and the UK in person, and remotely with teams in Europe and the US.
Do clients own the code BrashPixels writes?
Yes, entirely. All code, prompts, eval suites and documentation are delivered into the client's own repositories and owned outright by the client. There is no licence, no proprietary layer and no retainer required to keep anything running.
Get in touch
Tell us what you are trying to build and where you have got stuck. Straight answer, usually within a working day — including if the answer is that you do not need us. Or email hello@brashpixels.com.