Skip to content
dhii.ai

AI Infrastructure & MLOps

Private & Self-Hosted LLMs

ChatGPT-class assistants and APIs running on infrastructure you control, so prompts, documents and outputs never leave your environment.

Pilot deployment
1–2 weeks
Production with SSO
3–6 weeks
Typical team
1–2 engineers

When data can’t leave the building

For many organisations the blocker to using AI isn’t capability — it’s that contracts, regulation or common sense forbid sending client files, source code, patient notes or financial records to a third-party API. Self-hosting open-weight models removes that blocker: the model runs on your hardware or in your private cloud account, and nothing is sent elsewhere.

We run this stack ourselves. Dhii Chat is a self-hosted AI workspace built on Open WebUI and Ollama, operated on our own server. The lessons from running it — authentication, mail flows for sign-up and password reset, branding, model management, updates — go into every client deployment.

Picking the right model

There is no single best open model. We benchmark candidates on your tasks — your documents, your questions, your languages — and report quality alongside speed, memory footprint and licence terms (some open models have usage restrictions). Quantisation often lets a larger, better model fit on smaller hardware with minimal quality loss; we measure that trade-off rather than assuming it.

Production-grade serving

For a handful of users, Ollama on a single machine is enough. For a company-wide assistant or an API used by several applications, we deploy vLLM or TGI with continuous batching, behind an OpenAI-compatible gateway such as LiteLLM so your apps can switch between local and hosted models with a configuration change.

A usable workspace

People need an interface, not an endpoint. Open WebUI provides a polished chat experience with document upload, shared prompts, user groups and admin controls; we integrate it with your single sign-on and set sensible usage limits. Where you need something more tailored, we build a custom interface on the same backend.

Everything is documented: how to add a model, how to upgrade safely, how to restore from backup, and what to check when something is slow.

Deliverables

What you get

  • 01

    Model selection

    Benchmark of open-weight models on your tasks, with quality, speed, memory and licence notes.

  • 02

    Hardware plan

    GPU or CPU sizing for your users and workloads, on-premises or in your cloud account.

  • 03

    Inference server

    vLLM, TGI, llama.cpp or Ollama with an OpenAI-compatible API, batching and quantisation.

  • 04

    Chat workspace

    Open WebUI or a custom UI with SSO, user groups, document chat and usage limits.

  • 05

    Security hardening

    Network isolation, TLS, authentication, logging and update procedures.

  • 06

    Operations runbook

    Monitoring, backups, model upgrades and incident steps.

Process

How a private & self-hosted llms project runs

  1. Step 1

    Requirements

    Users, workloads, data sensitivity, latency targets and budget.

  2. Step 2

    Benchmark

    Test candidate models on your real tasks and hardware options.

  3. Step 3

    Deploy

    Provision infrastructure, inference server, UI and authentication.

  4. Step 4

    Harden

    Lock down access, add monitoring and run load tests.

  5. Step 5

    Hand over

    Train admins and users, and document upgrades.

Engagement models

Ways to work together

Every engagement starts with a fixed, written quote. See engagement models & pricing or the how much does llm fine-tuning cost?.

FAQ

Private & Self-Hosted LLMs: common questions

Are open-weight models good enough compared with GPT or Claude?

For many business tasks — drafting, summarising, extraction, classification, RAG over internal documents — current open-weight models are very capable. Frontier APIs still lead on the hardest reasoning and coding tasks. A hybrid setup that keeps sensitive work local and routes non-sensitive hard tasks to an API is common.

What hardware do we need?

It depends on model size and concurrent users. Small models run on a single modern GPU or even CPU; larger models need one or more high-memory GPUs. We size it from your workload and test before you buy or rent.

Ollama or vLLM?

Ollama is simple and great for single-server and low-concurrency use. vLLM delivers much higher throughput for many simultaneous users and is our usual choice for production APIs.

Do you run your own self-hosted AI?

Yes. Dhii Chat at chat.dhii.ai runs Open WebUI with locally served models via Ollama — the same building blocks we deploy for clients.

Related services

Start a project

Have a workflow that should run itself?

Tell us what you want automated or built. You get a written assessment of feasibility, risks and a fixed quote — usually within two business days.