What I Build

For the everyday 90% (drafting, summarising, answering questions from your own documents) you will not tell the difference. The current generation of open models is genuinely good, and it runs on hardware a small business can afford. For frontier reasoning on hard novel problems, the big APIs are still ahead, and I will say so.

Less than you think. A single well-specified workstation covers most small teams, and you likely have something close already. Sizing it for your real workload rather than a benchmark is part of the audit. I would rather tell you a £1,500 machine is enough than sell you a rack.

This is the strongest argument for running locally. If nothing leaves your building, there is no third-party processor to assess, no data-transfer agreement to sign, and no vendor whose retention policy you have to trust. For anyone handling client records, contracts, or health data, that is usually the whole conversation.

Your data never leaves. That is the exact claim, and it is worth being precise about, because the model itself has to arrive from somewhere. Pulling a model is a one-off download from a registry, over the network, at a moment you choose, and it happens before any of your documents are near it. After that, every prompt and every answer stays on the machine: no API call, no per-query egress, no third-party processor to assess. Compare that with a hosted API, where every single query leaves the building, permanently.

By default Ollama listens without authentication and phones home to check for new versions. Neither is acceptable on a network that takes itself seriously, so hardening is part of the setup, not an extra: the API is bound to localhost and put behind auth rather than left on 0.0.0.0 for the whole office, automatic update checks are turned off so the box makes no outbound call you did not ask for, and models are pre-pulled. If you need it genuinely air-gapped, it can be, and after the models are on the machine it never needs to see the internet again.

Often you would not, and I will say so. If Copilot is answering your questions well over the documents you keep in Microsoft 365, keep it: you are already paying for it. Where it stops is when the answer has to be grounded in a specific corpus with citations you can audit, when the per-seat bill scales faster than the value, or when the data genuinely cannot go to anyone else’s cloud on anyone’s terms. That is the gap I build for, and the audit exists to tell you honestly which side of it you are on.

When your volume is genuinely low, the API bill is not hurting, and your data is not sensitive. In that case you are paying me to save you money you were not really spending. Every engagement starts with an honest audit, and sometimes it concludes “stay on the API”. You get that in writing too.

Expected, and priced in. Every deliverable includes two rounds of revisions, where a round is one consolidated set of changes rather than a trickle of one-liners. After that, further rounds are £1,275 each, fixed, so you can weigh up whether a change is worth it before you ask for it. Anything that is genuinely new scope rather than a revision gets quoted before I start, never after.

You own everything. Open models, your hardware, your data, and handover documentation written for whoever comes after me. Nothing about a local setup depends on me still being around, which is rather the point of owning it rather than renting it.