Skip to content
Brad Stancel

Private AI: self-hosted LLMs that keep your data in-house

I set up AI that runs on hardware you control or behind a gateway you own, so your team gets modern tools without handing your data to someone else.

I want AI working on client files, contracts, and financials, but that data can't go into a public chatbot.

Public AI tools are useful right up until the work involves something confidential. Then the choice becomes either not using AI, or quietly using it anyway and hoping nothing sensitive leaks.

There's a third option. I run language models on my own hardware every day, including a two-node NVIDIA DGX Spark cluster, and I build the same kind of setup for businesses: models you host, a single gateway that controls who can use what, and tools that let AI read your systems without copying your data somewhere else.

How the work goes

  1. Classify the data

    We decide what can go to cloud models, what must stay local, and what AI shouldn't touch at all.

  2. Size the hardware honestly

    Plenty of useful work runs on modest hardware. I'll tell you when you don't need a GPU cluster, and when a cloud model with a proper agreement is the better call.

  3. Build the gateway and tools

    One front door for every model, with the tool connections your team actually needs.

  4. Prove it with your real work

    We test against your documents and questions, measure answer quality, and fix what the tests reveal before rollout.

Where I've done this

  • Operate a two-node NVIDIA DGX Spark cluster for local models. Built NCCL from source to fix a deadlock on the new GB10 chips, for 2.4× the throughput. See more

  • Run 25-plus AI tool servers behind a single self-hosted API gateway, used by coding agents and chat interfaces alike. See more

  • Built a fully local voice assistant for my home: speech recognition, a local language model, and speech synthesis, with no cloud dependency. See more

Common questions

Are local models good enough?

For many business tasks, yes: summarizing, extracting fields, drafting, answering questions from your documents. For the hardest reasoning, a cloud model may still win, and a gateway lets you use each where it fits.

What does the hardware cost?

It ranges from a single workstation to dedicated AI servers. We size it from your actual workload, and sometimes the answer is to start in the cloud with the right data agreement.

Can my existing software use it?

Usually. Most private model servers speak the same API as the major cloud providers, so tools that support those APIs can point at your own server instead.

Tell me what's slowing the business down.

Describe the problem in plain terms. I read every message myself and reply when there's something I can genuinely help with.

Start a conversation