Loading post
Notes from the workshop.
Writing about web development, AI tooling, and the occasional lesson learned the hard way.
Loading posts
Notes from the workshop.
Writing about web development, AI tooling, and the occasional lesson learned the hard way.
Loading posts
Loading post
Tell us what you're building. We'll tell you if we're the right fit — and if not, we'll point you somewhere that is.
Some of our AI runs on our own hardware, in the studio, not in someone else's cloud. With hosted models as good as they are now, it's a fair question why we bother.
The short answer: three reasons, and none of them is "local is always better". It isn't.
Client documents, early drafts, anything that shouldn't be uploaded anywhere: on our own hardware, it never leaves the building. That isn't a policy we promise to follow. It's where the files physically are.
That matters most for the material people are least comfortable sending away. Enquiries are a good example. Someone writes to us with their name, their business, what's gone wrong and what they can spend. That's exactly the kind of thing that should stay in the building.
Good work takes a lot of drafts, and most of them get thrown away. When every attempt is billed, there's a quiet pull to settle for the version that's nearly right.
Here, a draft gets redone because it should be, not because a usage line allows it. It isn't free. The hardware and the hours are real, and they're priced into the work. But nothing is counting tries.
Our line-up isn't fixed. Document work ran on Nemotron until Google's quantization-aware Gemma 4 releases did the same job on less hardware. Switching meant a download, not a renegotiation.
Gemma 4 12B is a good example of why that matters. It was trained with compression built in rather than squeezed afterwards, so it keeps close to full quality at roughly a quarter of the memory. It's now the model we reach for most often.
In our Lab there's a small working example. A rambling enquiry email goes in, and a structured brief comes out: who they are, what they need, budget, deadline, best time to call. The output is checked against a schema before anything uses it, so a model that wanders off script gets caught rather than passed along.
The steps are always the same: the input, a local model, a schema check, then structured output. The model in the middle can change. The check doesn't.
Each one has a job:
They run in LM Studio and Ollama. Search by meaning, rather than by keyword, runs locally too.
Claude is in daily use across our own work, with Gemini as the second option. Haiku, Sonnet, Opus and Fable each suit a different size of task, and knowing which to reach for is most of the skill.
Local and hosted are two tools, not two camps. What settles it, job by job, is two questions:
For client work, the choice and the reasoning are written down, so you know where your material went and why.
If you're weighing up AI for your business and want to know which side of that line your work falls on, start a brief or see the full line-up in the Lab.