Custom LLMs on open models, tuned on your data and run on your own servers, so nothing leaves your network.
The six setups we deliver most often, and any other your data rules call for.
Open models running on your own servers, with no route to the internet where your rules demand it.
Models served from your own cloud account, in the region your data has to stay in.
A private model for sensitive and high-volume work, and a hosted frontier model only for what it does better.
Open models adapted to your documents, terms and formats, and smaller models trained for a single job.
Document search, assistants and AI agents built on the private model, so your data stays in your network end to end.
The model kept fast, affordable and up to date after launch, with every request accounted for.
Six steps, from what the model has to do to a model your team uses every day on your own hardware.
Your use cases, data and compliance, and what it costs against a hosted API.
Open models tested on your cases and one chosen, with the hardware sized on-premises or in a private cloud.
Fine-tuned or given retrieval where it pays, and connected to your apps.
Accuracy, latency and hallucinations measured, with access and audit controls checked.
Served behind your own gateway, with single sign-on and monitoring.
Model updates, patches and on-call, or a handover to your own team.
Your use cases, data and compliance, and what it costs against a hosted API.
Open models tested on your cases and one chosen, with the hardware sized on-premises or in a private cloud.
Fine-tuned or given retrieval where it pays, and connected to your apps.
Accuracy, latency and hallucinations measured, with access and audit controls checked.
Served behind your own gateway, with single sign-on and monitoring.
Model updates, patches and on-call, or a handover to your own team.
What we build into a private model, set for where it runs and who uses it.
No calls to outside AI services: prompts, files and answers stay on infrastructure you control.
An OpenAI-compatible endpoint, so existing tools and code switch over by changing one address.
People sign in with your single sign-on, and each team reaches only the models and data it should.
Who asked what, when, and which model answered, kept in a log you own.
Quotas on requests and tokens for each team or app, so one workload can't starve the rest.
Models quantised and batched to fit the GPUs you have, without falling below the quality bar.
Each new open model scored on your cases, and swapped in only when it wins.
Throughput, latency, errors and GPU load tracked, with alerts before your users notice.
Prompt injection, data leaks and access rules tested before launch and after every change.
Starting something new with AI, or adding it to software you already have.
An AI MVP with the features that matter most, to launch and learn.
An AI product designed, built and launched in full, by one team.
AI engineers join your team, or a dedicated team runs the work.
AI built into the systems you already run, working with your data and permissions.
A demo or proof of concept made reliable, secure and measured, ready for real users.
Hosted and monitored, with models evaluated and updated as your data changes, agreed with you.
Custom software development across industries: web, mobile and AI applications, built and security-audited by one team.
We also build software for insurance (insurtech), government (govtech), energy and media. If your industry is not listed, we can still build for it.
Three products of our own, designed, built and kept running by the team that would deploy your private LLM.
Retail · SaaS
A point-of-sale system that keeps sales, stock and receipts together on till, tablet and phone.
Travel
A travel booking app whose AI assistant finds flights, hotels and trip ideas, through to a confirmed booking.
AI
An AI assistant for customer support and sales that answers questions and guides buyers.
What teams ask before they run a model on their own infrastructure, answered plainly.
When your data can't go to an outside provider, when your volume makes per-token pricing expensive, or when you need control over speed and uptime. For a prototype, or for the hardest reasoning, a hosted API is often the better start, and we'll say so. Many teams end up hybrid: private for sensitive and routine work, hosted for the rest.
Yes. We deploy on-premise, in your own cloud account or fully air-gapped. For an air-gapped install, the models, their dependencies and every update are brought in offline, and the servers have no route out.
Llama, Qwen, Mistral, Gemma and DeepSeek, among others. The best open models now match hosted ones on many everyday tasks, and a fine-tuned model can beat them on yours — but not on everything. We test candidates on your own examples first, so you see the gap before you decide.
It depends on the model's size and your traffic. A small model runs on a single GPU; a large one needs several, or a quantised version that fits on fewer. We size the hardware to your workload before anything is bought, and can work with servers you already have.
For answering from your documents, RAG is usually enough, and it keeps facts current. Fine-tuning suits a format, a tone or a narrow task, and lets a smaller model do a bigger one's job. Training a model from scratch is rarely worth it; we'll tell you if your case is the exception.
Usually a few hundred to a few thousand good examples of the task, which we help you collect and clean. Training runs on your infrastructure or in your cloud account, so the data never leaves it, and personal data can be removed first.
Yes. The model serves an OpenAI-compatible API, so most tools and code connect as they are. On top of it we build private RAG over your documents, assistants for your teams and AI agents that act in your systems.
Prompts, files and answers stay inside your network. People sign in through your single sign-on, each team reaches only what it should, every request is logged, and the deployment is tested for prompt injection and data leaks before launch.
Usually you do: for client work, the fine-tuned weights, the training data, the code and the configuration are yours, as set out in our agreement, and run on infrastructure you control. Open model licences differ, and we check that each one fits your use.
Testing models on your cases and sizing the hardware takes one to two weeks, a first deployment two to six, and fine-tuning a few weeks more where it pays. Running cost is mostly the hardware or cloud GPUs, a fixed cost rather than a bill per token. After a first call we send a written estimate.
