Local AI Pilot
One AI feature, running on your own GPU, in 30 days.
For software teams that want AI in their product but do not want a per-request bill, or cannot send their data to a cloud model. I build one working feature on hardware you control, measured against a cloud model on your own data, and I hand it over to your team.
The problem
A cloud model is easy to try and hard to live with. The bill grows with every user, a stuck loop can cost thousands overnight, and some data is not allowed to leave your servers at all. Small open models can do many of these tasks well on a modest graphics card, but only if someone measures which model, on which card, for which task, and builds the checks around it.
How it works
1. Scoping week
We pick one task worth automating, collect a test set from your real data, and choose the hardware: your own GPU or a rented one with a fixed monthly cap. At the end of the week you get a written go / no-go with numbers.
2. Baseline against a local model
I run the same test set through a cloud model (Claude or GPT) and through small open models that fit your card. You see accuracy, speed and cost side by side before anything is built.
3. Build with Claude Code
The working feature: a chain of small models with checks between them, a job queue so the GPU is never overloaded, and a cloud model only for the cases the local one is unsure about. Every change is measured against the test set.
4. Hand-over
A runbook, a page that shows the test results after every change, and a working session with your developers on how I use Claude Code, so your team can carry it on without me.
What you get
- No bill that grows with traffic. The model runs on a card you own, or one you rent at a fixed price.
- Your data stays on your servers. Nothing is sent to a cloud model unless you decide it should be.
- Numbers, not opinions. You know how accurate it is, how fast it is and what it costs before you commit to more.
Proof from my own projects
- A small fine-tuned model sorted banking customer messages into 77 intents with 94.2% accuracy in English, against 85.3% for Claude Sonnet, in 12 ms instead of 2 seconds. In Hebrew: 82.8% against 77.3%.
- Seferehov reads Hebrew book spines from a phone photo. Adding a Hebrew OCR model and a spell-check took text accuracy from 65% to 96%, on one 8 GB card.
- Kartis7, a ticket-sales platform, checks every new event page for scams with a model on that same card.
- For a compliance software company: AI that runs on the client's own office GPUs, with a fallback to CPU when the GPU is busy, so client evidence never leaves their country.
Price
- Local AI Pilot, 30 days: $12,000, fixed.
- Scoping week on its own: $4,500, credited in full if you continue to the Pilot. If the week does not end with a clear go / no-go backed by numbers, I refund it.
- Pilot plus three months of advice: $12,000 and then $3,500 a month for a weekly review with your team.
- Not ready yet? A two-hour review of your GPU needs and your team's Claude Code workflow, with a written report: $1,000.
Questions
What hardware do I need?
Often less than you think. My own projects run on one 2016 graphics card with 8 GB. If you have no GPU, we rent one by the month, and the scoping week tells you which size.
What kind of tasks fit?
Tasks with a clear right answer that your team does many times a day: sorting customer messages, reading documents or images, pulling fields out of forms, searching your own documents. Open-ended writing is usually better left to a cloud model.
Do you work with our developers or instead of them?
With them. I build the first version, and the hand-over is designed so your team owns the code from day one.
What if the scoping week shows a local model is not good enough?
Then I tell you that, with the numbers, and you decide whether to stop there. If the week does not give you a clear go / no-go, I refund it.
Tell me the task you have in mind. In 30 minutes I will tell you whether a local model can do it, roughly which card it needs, and whether the Pilot is the right fit.
Book a scoping call