Paddock
Your manuals, policies, and knowledge, made answerable with exact citations and table-precise answers. It runs on your hardware and nothing leaves the fence.
Explore Paddock
About Black Sheep AI
Frontier intelligence you fully own
Black Sheep AI is a research and deployment firm. We do original work in quantization, training, and distillation, and we engineer all of it to run on infrastructure you control. Your models stay on your hardware. Your data never leaves the fence.
Why We Exist
The conviction behind the work
We started Black Sheep AI on a plain observation. The best models kept getting larger and further out of reach, locked behind someone else's API on someone else's servers. If you cared where your data went, you were out of options.
We think that is backwards. A hospital, a bank, a defence agency, a firm with fifty years of hard-won documents, none of them should have to hand their crown jewels to a third party to get useful AI. So we set out to make frontier models small enough to run on hardware you already own, and honest enough that you can prove what they do.
The work is research first. We compress models with 400B+ parameters down to a size that fits on a single GPU instance or a Mac Studio, and the compressed model holds the full-precision scores it started with. Then we ship it, on premises, air-gapped, or at the edge, wherever your data has to stay.
We are a team of researchers and deployment engineers working from Australia and New Zealand. We help nations and enterprises build AI capability they own outright, from the weights on disk to the box that runs them.
How We Work
Principles we don't bend
The same rules sit behind every model we compress and every system we deploy.
01
Original research, not wrappers
Every capability we offer comes out of our own research. We don't put a logo on someone else's model and call it a product. We invent the method, test it, and deploy it.
02
Evidence over claims
We publish our methods and our numbers. When we say a compressed model matches its full-precision score, there is a report with the measurement behind it. Over twenty articles are already out in the open.
03
Runs on hardware you control
If it needs a cloud we operate, it isn't finished. Everything we build targets infrastructure you control, down to air-gapped rooms with no network at all.
04
Honest about limitations
Every report states accuracy and its limits in full. We tell you what a model can't do as plainly as what it can. Research that only reports its wins is marketing.
The Product Family
What we put in your hands
Three products, each doing one job well, all of them running behind your own fence.
Watchman
Verify a model is what it claims to be before you deploy. It detects and classifies weight modifications in third-party and compressed releases, with evidence-grade reports and AI-BOM attestations.
Audit before you deployShepherd
Production-grade model compression with CI/CD integration and fleet deployment. Every build is certified by Watchman before it ships.
See Shepherd
How We Think
The frameworks behind the work
Two bodies of thinking shape what we build and how we tell customers to deploy it. We publish both so you can argue with them.
The RAI Framework
How we make responsible AI a set of checks a team can actually run, mapped to the regulation you answer to, with Watchman producing the evidence. Our position on what accountable deployment looks like in practice.
Read the frameworkThe AI Innovation Framework
Our model for taking an AI idea from a claim to a deployed, verified system without the usual waste. The way we run our own lab, written down.
Read the frameworkRun private AI on hardware you control
Original research, production engineering, and sovereign deployment, from one team. Tell us what has to stay on your side of the fence.
Talk to Our Team