Serving
Open-weight models on your hardware.
Private AI Foundation
Installed once. Applications attach.
The runtime is the product. Model serving, Model Gateway, governance, RBAC, immutable audit, data protection and offline diagnostics. DocxIntel and AgentLab sit on it. They do not replace it.
Gateway
One path for prompts and tools.
Controls
Policy, RBAC and audit in the path.
Offline
Diagnostics as a file you inspect.
What the Foundation does
Install the runtime
One environment. Air-gapped, tenancy or bare metal.
Load the weights
Apache 2.0 and MIT files on disk, with licence text beside them.
Attach applications
DocxIntel and AgentLab use the same serving, gateway and audit.
Operate
Updates as signed offline bundles. Nothing calls home.
Runtime capabilities
Model serving
Vision/OCR, LLM, embeddings, reranking, four classes, open licences.
Model Gateway
Every prompt and tool call addressed through one path.
Governance
Policy, DLP and HITL live in the path of inference, not after it.
RBAC
Named access to model classes and data classes.
Immutable audit
Written where the workload runs. Survives a pulled cable.
Offline diagnostics
A file you inspect before it goes anywhere. Not a phone-home agent.
Inside the install
Serving
- Vision / OCR
- LLM
- Embeddings
- Reranking
Four capability classes. Apache 2.0 and MIT only.
Control plane
- Model Gateway
- Policy
- RBAC
- Guardrails
Who may ask what of which model, on which data.
Evidence
- Immutable logs
- Audit fields
- Licence files on disk
- SBOM per release
Your legal team reads files, not assurances.
Protection
- Customer-held keys
- No BizfyLabs endpoint
- No training on your documents
- Offline bundles
Nothing you process is retained by us, because it never reaches us.
Why start here
Installed once. Applications attach.
If you start with a parser
You have bought a feature. If you start with the runtime, you have bought a perimeter.
Same review, two applications
DocxIntel and AgentLab inherit the Foundation. One security pack.
Hardware is not a gate
Starter is 8 vCPU and 32 GB. A GPU is never a precondition for beginning.
You hold the keys
In every tier that touches customer data. That does not change with the substrate.
The full stack on one machine, disconnected from the network.
In your office, on your documents, with the cable pulled out. No competitor's sales engineer can do this.