What I take on
Private AI & Custom Models.
Built by me, start to deploy
Your own model, on your own hardware, with nothing leaving the building.
Why this gets built
Two things stop most companies putting AI near their real data. The data cannot go to a third party, and the per-user pricing of the tools that could handle it does not survive contact with the headcount. A model you host yourself answers both, and on a narrow task it is often the better model anyway.
What I build
- A small model tuned to your task instead of a large general one rented by the token. Narrow work is where this wins, and it is also the cheapest place to prove it.
- Quantization and GPU inference sized to your actual load, so the hardware bill is a number you can check rather than an invoice that moves every month.
- Retrieval over your own documents with your permissions carried through, so a person only gets answers from what they were already allowed to read.
- An eval set drawn from your real cases, run before anything is switched over, because a private model is only worth it if the answers hold.
- Deployment inside your own network or cloud account. I do not need a copy of your data and I would rather not have one.
- An optional fallback to a hosted model for the cases the local one should not take, with a switch that turns it off entirely.
How long it takes
Around four weeks to a private model answering on your own hardware, once there is one task and a set of real cases to test it against.
This starts on a call, not with a quote. What it costs depends on the task, the volume and the hardware you already have, and half an hour is usually enough for me to say whether a private model beats the API for you, or whether it does not and you should stay where you are.
What ships
- The tuned model and its weights, yours
- Inference deployment in your own environment
- Retrieval over your documents with permissions
- Eval results against your real cases
- Running cost per thousand requests
Tech stack
Next step
Let’s talk, 30 minutesNeed this built?
Tell me what you’re running and where it breaks. Half an hour is usually enough for me to say what it takes and what I’d do first.