AI / ML · In depth
Your hardware. Your models. Your data.
We design and deploy AI that runs on machines you control. Local open-weight models, no third-party APIs, so sensitive data stays in your environment and costs stay predictable.
What does on-premise AI actually mean?
It means the model runs on hardware you control instead of behind someone else's API. Open-weight models like Llama and Gemma are now good enough for most production tasks that are not frontier reasoning, and runtimes like Ollama and llama.cpp make serving them a matter of one command. Quantization, running a model at 4-bit instead of 16-bit, cuts its memory footprint by roughly four times, so a 27-billion-parameter model fits on a single prosumer GPU or a Mac with enough unified memory. A mid-range machine handles classification, extraction, and drafting at interactive speed. That is not a lab setup. That is a desk.
It does not have to mean a server room, either. Some models run on hardware you own; others can run on private infrastructure or cloud GPUs in accounts you control. Either way, your data never flows through a third-party AI API.
What problems does it solve?
Your data stays in your environment
The moment a customer record, a contract, or a medical note leaves your network for an external API, you have created a compliance surface you must answer for. On-premise inference removes the question: nothing leaves. For work under the GDPR, and for clients who do not want their data training someone else's next model, this is not a feature. It is the requirement.
Costs become fixed and knowable
Per-token pricing is fine until volume grows, and then it is a tax that scales with your success. Local inference costs what the hardware and the electricity cost. Past a certain usage, owning the compute is cheaper, and the cost curve is flat instead of climbing.
No vendor can pull the floor out
Hosted APIs deprecate models, change pricing, rate limit you at the worst moment, and occasionally go down. A model on your hardware behaves the same today as it did six months ago, because it is the same weights on the same silicon. That reproducibility matters when you are building something you intend to keep.
What are the honest trade-offs?
Frontier hosted models are still ahead on the hardest reasoning, the longest context, and the most obscure knowledge. If your product lives or dies on that top slice of capability, an API is the right tool, and we will say so.
You also take on the hardware. Something has to run, stay patched, and be monitored. That is real operational work, though far less than most teams fear once it is set up properly.
And smaller local models reward good engineering. A 7B model with clean retrieval, tight prompts, and a narrow task will outperform a giant model pointed at a vague one. The work moves from paying for a bigger model to designing the system well, which is the work we prefer anyway.
Where does it already run?
Our own lab is the standing proof: four machines on a private Tailscale mesh, with no public IPs, running Gemma and Llama-class models through Ollama on Apple Silicon, plus a GPU desktop that wakes only for heavy rendering. The active nodes idle in the single-to-low-double-digit watts.
See the hardwareIt ships in real projects too. A recent build serves a real-time 3D site in five languages, every one translated by a model running locally through Ollama, at zero API cost.
Read the case studyWhen is it the right call?
A simple test. On-premise AI tends to win when at least one of these is true:
- The data is sensitive, regulated, or contractually cannot leave your environment.
- The volume is high enough that per-token cost is a real line item.
- You need the behaviour to be stable and reproducible over time.
- The task is well defined (extraction, classification, search, translation, drafting, routing) rather than open-ended frontier reasoning.
If none of those hold, use an API and move on; we will tell you that in the first call. If one or more do, running it yourself is usually the calmer, cheaper, more defensible path, and it is the one we know how to build.
Where most clients start
The Sovereign AI Pilot
In two to three weeks, we take one real workflow from your business and make it run on AI you own. Fixed scope, one workflow, a clear answer.
Your data never leaves your control, from day one of the pilot.
Fixed scope · 2–3 weeks · The proposal names the number
A working prototype on your hardware, or on ours temporarily
A plain-language report of what worked and what it would cost to run for real
A go or no-go recommendation you can trust, including "do not do this" if that is the honest answer
FAQ
Common questions.
What hardware does on-premise AI need?
In practice: one machine sized to the model (a GPU box or Apple Silicon with enough unified memory), Ollama as the runtime, a small queue in front, and quantized open-weight models with headroom to spare. Retrieval over your own documents and a thin internal API complete the picture; the rest of your stack never needs to know where the model lives.
Does on-premise mean we need our own server room?
No. The models can run on hardware you own, or on private infrastructure and cloud GPUs in accounts you control. Either way, your data never flows through a third-party AI API, which is the part that matters.
Who owns the system you build?
You do. The code, the models' configuration, the accounts, and the data are all handed over to you in full at launch, so the system never depends on us to keep running.