AI & automation
AI where there's a workflow to plug it into.
We don't sell models. We look at work you already do — research, publishing, data evaluation — and put automation where it removes a real step.
What it's good at, and what it isn't
Current models are good at a narrow set of things: drafting, extraction, classification, summarising and code. They are bad at facts you cannot check and judgement you cannot audit.
So we build with the check-step in, not bolted on. Something a person looks at, or a rule that fails loudly. If a workflow can't tolerate being wrong occasionally, it needs a different shape — and sometimes the honest answer is that it shouldn't be automated at all.
Four kinds of work
Where we have actually done this.
Science and research
Literature triage, structured extraction from papers and PDFs, dataset annotation, and pipelines you can run again and get the same answer.
Web and CMS
Drafting with review gates, translation, metadata and alt text at scale, internal search, and editorial assistants that work inside the CMS you already use.
Data evaluation and analysis
Normalising messy input, classification, reconciling sources that disagree, and drafting the report a person then signs.
Workflow plumbing
Assistants wired to your own tools, scheduled jobs, API integrations, and self-hosting where the data can't leave the building.
Your data, your rules
Most of the risk in this work isn’t the model being wrong. It’s not knowing where your data went, who can see it, or what you agreed to.
So that gets decided first and written down, before anything is built.
We name the provider
Which model, which tier, in the contract. Not “we use AI”.
Self-hosted is a real option
Slower and cheaper to run, and nothing leaves your infrastructure.
Logged and reproducible
Inputs, outputs and prompts kept, so you can answer questions about a decision months later.
How an engagement is shaped
- 01
Workflow audit
A fixed fee. We watch how the work is actually done, then write down where automation would remove a step and where it would only add one.
- 02
Prototype in two weeks
One workflow, end to end, on your real data. Enough to know whether it's worth building properly.
- 03
Build and hand over
Documented, self-hostable, yours. We'd rather you didn't need us afterwards.
- 04
Retainer for the watching
Only for the parts that genuinely need watching — model changes, cost drift, quality regressions. Cancellable.
The questions that matter
Which models do you use?
Whichever fits the job and the constraints, and we name it in the contract. For work that can't leave your infrastructure, an open-weights model you host yourself.
Is our data used for training?
We use provider tiers whose terms exclude training on customer data, and we name the provider and the tier in your contract so it's a commitment rather than a reassurance. If that still isn't good enough for your data, we build it self-hosted and nothing leaves your network.
Can it run on-premise?
Yes, and for some kinds of data it's the only version we'll build. Expect it to be slower and to need a GPU you either own or rent.
How do you stop it inventing things?
You don't stop it — you catch it. Structured output with validation, retrieval against your own sources with citations, and a human check-step where being wrong would matter. Anything that claims to eliminate the problem is selling you something.
What does it cost to run per month?
We estimate it during the prototype, with the real token counts from your real data, and design to a ceiling you set. Runaway cost is a design failure, not a surprise.
Can you just make our chatbot better?
Sometimes, and often the honest answer is that the chatbot is the wrong shape for the problem. We'll say which.
Start with the audit.
It is a fixed fee and it sometimes concludes that you should not build anything. That is a useful outcome too.