Private AI, built where your data lives.

Services

What we do

• Private & on-prem LLMs

  Open models running on your own servers, your own cloud account, or a box with no internet cable at all. Prompts and documents stay on your network, and nobody else’s terms of service get a say in what happens to them.

• LLM fine-tuning

  Teach a good open model your vocabulary, your formats and the way you write. A small tuned model can beat a big general one at a narrow job, for a fraction of the running cost. When it can’t, we’ll say so.

• RAG over your documents

  Ask questions of your policies, manuals and tickets and get answers that point to the page they came from. If the answer isn’t in your documents, the system should say so rather than make one up, and we test that it does.

• MLOps & deployment

  Pipelines you can re-run next year and get the same result. Kubernetes where it earns its keep, monitoring that tells you when quality drifts, and a GPU bill you can predict.

• AI evaluation & guardrails

  Before launch, we test against the questions your users will actually ask, awkward ones included. After launch, the same tests keep running, so you find out quality slipped before your customers do.

Projects

Example projects

The kind of work we take on. Each one ends with something running on your infrastructure, and the documentation to keep it there.

• Private LLM gateway

  One internal endpoint for every team’s AI use. Open models served on your own GPUs, single sign-on, per-team quotas, and every request logged where your security team can see it.

  • vLLM
  • Kubernetes
  • OIDC
  • Prometheus

• Document assistant with citations

  Questions over policies, contracts and tickets, with answers that link back to the paragraph they came from. Permissions follow the documents, so nobody gets answers from files they couldn’t open anyway.

  • RAG
  • pgvector
  • open embeddings
  • Python

• GPU platform on Kubernetes

  A small GPU cluster, on-prem or in your cloud account, with job queues, autoscaling and cost dashboards, so training and inference share the hardware without fighting over it.

  • Kubernetes
  • NVIDIA GPU Operator
  • Kueue
  • Grafana

• Cloud landing zone as code

  A cloud account set up properly from day one: networks, identity, encryption and logging written in Terraform, reviewed in pull requests, and rebuildable from scratch.

  • Terraform
  • OpenTofu
  • GitHub Actions

• Domain fine-tune

  A small open model trained on your own labelled data for one job, like sorting claims, drafting replies in your house style or pulling fields out of forms. Measured against a big general model before anyone switches.

  • LoRA
  • Hugging Face
  • eval harness

• Evaluation & red-team suite

  Automated tests for an AI system: accuracy on your real questions, refusing what it shouldn’t answer, personal-data leaks, and drift over time. Runs in CI and produces a report a compliance team can read.

  • Python
  • CI
  • test sets
  • reports

Who we help

Built for teams that need to stay in control

• SMBs & startups

  You want AI that does one useful job, not a pilot that never ends. We start small, keep running costs visible from day one, and build on open models and standard tools, so you can walk away with everything.

• Regulated & privacy-sensitive teams

  Healthcare, finance, insurance, public sector: places where data can’t be pasted into someone else’s chatbot. Data stays in your environment, access is logged, and evaluation results are written so an auditor can follow them.

How we work

Three steps, no black boxes

1. Assess

   One to two weeks looking at your data, your infrastructure and the problem you actually have. You get a written plan: what to build, what it will cost to run, and what we’d leave out.

2. Build

   Short cycles and something working early. You see progress every week as a demo, not a status report.

3. Hand over

   Code, configs and runbooks go into your repository, and your team knows how to run it. You own all of it. We’re around if you want us, and you’re not stuck if you don’t.

FAQ

Fair questions

Do we need our own GPUs?

Not always. Plenty of useful models run on a single modern GPU, and some small ones run fine on CPU. Renting is also fine, as long as the cloud account is yours.

Which models do you use?

Open-weight models, picked for the job. We compare a few on your data before recommending one. If a hosted API is honestly the better fit for you, we’ll tell you that too.

Can you work with what we already have?

Yes. Most of the work is fitting into your existing infrastructure, identity and security rules, not replacing them.

What does it cost?

It depends on the scope, which is why the assessment ends with a written estimate. You’ll know the running costs before anything gets built.