If you searched "ai shepherd" and landed here, this is the one. Shepherd is the deployment pipeline of the Bounded AI Architecture: it takes a frontier model, compresses it into a Ram reasoning model, certifies it, binds it to your Paddock, and lands all of it on hardware you control. Here's what it does, what problem it solves, and why the compression underneath it holds up.
Most "compression" tools give you a smaller file and a shrug about what happened to the model inside it. You quantize a 31B model down to something that fits your hardware, run a couple of prompts, and hope the reasoning survived. It usually didn't, not all of it, and you find out in production when a legal summarization task starts inventing clauses.
Shepherd is our answer to that. It is the deployment pipeline of the Bounded AI Architecture (BAA). It takes a frontier model, compresses it into a Ram reasoning model with the RAM method, certifies the build with Watchman, binds it to your Paddock, and deploys the whole architecture to your infrastructure. Compression, certification and deployment in one pipeline, not three tools duct-taped together.
First, the name
"ai shepherd" is a search people type when they've heard the name and want the actual product. So, to be unambiguous: Shepherd is Black Sheep AI's deployment pipeline for Bounded AI. It's not a chatbot, not a prompt manager, not an agent framework. It takes the frontier model you pick, compresses it into a Ram reasoning model, certifies it with Watchman, binds it to your Paddock, and deploys all three layers privately: in your VPC, on an Apple Silicon fleet, or in an air-gapped facility. If that's what you were looking for, you're in the right place.
The problem Shepherd solves
Running a frontier model privately used to mean one of two bad options. Either you paid for multi-GPU infrastructure to hold the full-precision weights, or you sent your data to an API provider and gave up on privacy. Both are expensive, just in different currencies.
RAM compression, the method behind the Ram reasoning layer, changes the arithmetic. It shrinks frontier models by 50 to 60 percent, which cuts your hardware requirement roughly in half while keeping the intelligence you deployed the model for in the first place. A 31B-parameter model that needed workstation-class hardware at full precision now fits on a single GPU instance, or on a Mac Studio sitting at someone's desk. Your data never leaves your infrastructure, because the model runs on hardware you control.
That's the claim. The reason to believe it is the benchmark.
The proof: Gemma 4 at half the size
We compressed Gemma 4 31B to 31GB with the RAM method and ran it against the complete 12,032-question MMLU-Pro suite, the most demanding multiple-choice academic benchmark in wide use. It scored 85.2 percent, matching Google's published BF16 baseline exactly, on the full suite with no subset selection and no scoring shortcuts.
The full-precision version weighs roughly 62GB. Ours is 31GB. Same published number, half the footprint. Six categories, including math at 94.4%, biology at 92.7%, and physics at 89.8%, cleared 89 percent, the reasoning-heavy categories where broken quantization usually collapses. There were no catastrophic drops. The lowest category, law, tracked a known Gemma-family soft spot rather than a compression artifact.
We wrote the whole result up, per-category tables and all, in our Gemma 4 research report. It's the single best answer to the question "does your compression actually preserve capability."
What Shepherd deploys
Shepherd lands the three layers of the architecture on hardware you control. What arrives:
- Paddock, the Knowledge Plane. Your knowledge in Data Modules, each permissioned on its own. The model sees only what its modules allow.
- The bound Ram reasoning model. The frontier model you chose, shrunk to fit your GPUs with its reasoning kept and certified, relying strictly on your Data Modules for knowledge. When a better model ships, swap it; the plane does not change.
- Watchman governance. One log for every module permission, freeze, hand-back and deletion, plus provenance verification and certification of the bound model.
And the pipeline that gets them there and keeps them current:
- RAM compression, in production now. A data-free, post-training method built for both NVIDIA GPU and Apple Silicon targets. No calibration dataset required.
- CI/CD integration. Automated compression triggered on model release, with webhook triggers, CLI tooling, and API access. Compress once, deploy everywhere.
- Quality-gated releases. Every build is certified by Watchman before it ships. Models that miss your policy never reach production.
- Fleet and cluster rollout. Across Apple Silicon clusters, on-prem GPU infrastructure, or air-gapped and sovereign environments.
- Broad architecture support. Dense, Mixture-of-Experts, and multimodal models, on the same automated workflow.
- Batch compression for whole model catalogues, and multi-format export to MLX, GGUF, and custom targets.
- Dedicated engineering support and an SLA.
The hardware story is the part finance cares about. A 31GB Ram reasoning model fits comfortably on an AWS g7e.2xlarge (RTX PRO 6000, 96GB VRAM), an Azure NC24ads A100 80GB, or a GCP a2-highgpu-1g (A100 40GB).
| Target | Hardware | Indicative cost |
|---|---|---|
| AWS g7e.2xlarge | RTX PRO 6000, 96GB VRAM | ~$8,700/yr on-demand, ~$2,700/yr spot* |
| Azure NC24ads A100 v4 | A100 80GB | ~$9,500/yr on-demand, ~$2,900/yr low-priority* |
| GCP a2-highgpu-1g | A100 40GB | ~$9,500/yr on-demand, ~$2,900/yr preemptible* |
| Mac Studio | M4 Ultra, 192GB unified memory | ~$8,000 one-time, electricity after |
*List pricing before corporate cloud discounts, based on a 10-hour workday. The Mac Studio runs Ram reasoning models natively via MLX with no cloud dependency at all.
Compression you can hand to an auditor
Here's where Shepherd stops being a convenience tool and starts being something a regulated organization can actually deploy. Every Ram build Shepherd produces is certified by Watchman, the GRC layer of the architecture, before it goes live, and re-certified when it changes.
That certification is not a benchmark score. It's an evidence-grade provenance report plus a CycloneDX AI-BOM attestation confirming the compressed model is your declared base plus the declared quantization and nothing else, with pass/fail gating against your policy across five measured gates. Models that don't meet your standard never reach production. You get traceability from source model to deployed artifact, in one pipeline, with no separate tooling and no manual validation step.
This matters because compression is exactly the place a supply-chain problem can hide. A modified weight is easy to pass off as "just quantization noise" if nobody's checking. Watchman checks, on every build, which is why the GRC layer sits inside the Shepherd pipeline rather than beside it.
Where Shepherd fits
If you're weighing a private-AI deployment, the sequence Shepherd runs is the one you'd otherwise assemble by hand and get wrong somewhere: take the model, compress it into a Ram reasoning model without destroying capability, prove that's what you did, bind it to the knowledge it is allowed to see, and push it to hardware you control. The RAM benchmark says the capability survives. The Watchman certification says you can prove it. The Paddock bind says the model sees only what its modules allow. The cost tables say your finance team will sign off.
A free individual edition, rebuilt on the same RAM method, is coming for personal use on Apple Silicon. But the enterprise edition is available now, and it's the one that runs compression, certification, the bind and fleet rollout as a single pipeline.
