Notes on private AI, in practice.

What we learn shipping small models into real businesses — architecture decisions, deployment trade-offs and the things nobody writes in the docs.

Featured

Why a 350M model beats a frontier model for document extraction.

Accuracy on a narrow task is a function of training data, not parameter count. Here is what we measured across 40,000 invoices — and what it costs to run either option on your own hardware.

Read the article

Latest writing.

Engineering notes, deployment guides and field reports.

Architecture Latest Many small models beat one large one Routing each task to a specialist model cuts cost and raises accuracy. How we structure the router. Aug 2026|6 min read

Deployment Running inference on CPUs you already own Quantisation, batching and thread pinning — the three settings that decide whether CPU-only is viable. Jul 2026|8 min read

Security What “no data egress” actually requires Air-gapped does not mean offline. A checklist for proving your AI stack never phones home. Jul 2026|5 min read

Fine-tuning Building a training set from your own archive You probably have enough labelled documents already. Finding them is the work. Jun 2026|7 min read

Case study Invoice processing for a Dubai contractor 11,000 invoices a month, four languages, one CPU-only box in their own server room. Jun 2026|9 min read

Opinion The per-seat AI bill nobody budgeted for Why usage-based pricing breaks down the moment AI becomes part of everyday work. May 2026|4 min read

Beyond the writing

The work itself is more useful than the write-up.

Every note here comes out of a real deployment. If one of them describes your problem, the next step is a conversation or a test — not a reading list.

Try it Run your own documents Open the playground and put your messiest files through the current model. Scope it Tell us the task and hardware We will say honestly whether a small private model can do it — and if it cannot, we will say so. Read on How deployment works Every device it runs on, what it needs, and what hosting actually costs.