Control
AI on infrastructure you own
Models, retrieval indexes, prompts, and generated answers are designed to remain inside the customer-controlled environment.
Your AI. Your rack. Your rules.
A customer-site AI appliance with local models, governed document search, and an operating handoff your IT team can own.
Own the operating boundary
Useful AI, under your control.
Local models, governed search, and a complete operating handoff—inside the environment you own.
Access
Built for the way your team works
Office users connect over private HTTPS. Approved remote users connect through the customer’s VPN or zero-trust access path with MFA.
Operations
Delivered as a working system
Skilak stages, tests, transports, installs, and validates the appliance with your IT and security teams—not as a box left at the loading dock.
Where it sits · who touches it
One appliance. No public front door.
Local users connect over your network. Remote users come through your VPN or zero-trust path.
Authorized users
OfficePrivate LAN
RemoteManaged device
AdminPrivileged path
Private HTTPS · Remote access: MFA + VPN / ZTNA
Customer-controlled site
Approved network boundarySkilakSpoolLocked server room / secure IT space
01Private interface
02Model runtime
03Retrieval + data
Optional, approved egress only
IdentityUpdatesSupport
Right-sized systems
Start small. Scale after proof.
Choose a planning tier, then validate the final hardware against the work.
5–25 named users
Solo
A compact private-AI starting point for an executive team, proposal shop, engineering cell, or controlled pilot.
- Active use
- Typically 1–5 active generations
- Memory
- 48–128 GB accelerator memory; 96–192 GB unified-memory class where appropriate
- Platform
- Professional tower or compact rackmount appliance
- Model class
- Fast small-to-mid local models; larger quantized models after validation
- Private chat and document Q&A
- Optional retrieval with source citations
- Local user administration
- Acceptance testing and operator handoff
20–150 named users
Team
A rack-mounted deployment for shared departmental use, stronger identity integration, and sustained internal workloads.
- Active use
- Typically 6–24 active generations
- Memory
- 96–384 GB aggregate accelerator memory; 192–512 GB unified-memory class
- Platform
- Enterprise rack server or paired rack appliances
- Model class
- Mid-to-large local models selected for shared quality and throughput
- Microsoft Entra ID or compatible SSO
- Governed retrieval and citations
- Role-based access and audit logging
- Monitoring, backup, and restore runbooks
150–500+ named users
Enterprise
A site-specific architecture for multiple workloads, larger concurrency targets, or specialized model requirements.
- Active use
- 25–100+ active generations after workload validation
- Memory
- 384 GB–1.5 TB+ aggregate accelerator or unified memory, workload dependent
- Platform
- Multi-server rack architecture or segmented site deployment
- Model class
- A routed model portfolio selected per workflow, boundary, and service target
- Site survey and capacity planning
- Network and identity architecture
- High-availability options where justified
- Phased rollout and operational enablement
Prove the system
Acceptance is evidence, not a promise.
Each engagement defines what must work and what evidence will demonstrate it before handoff.
See the validation modelDisconnected rehearsal
The approved AI workflow remains available without a public internet path.
Grounded retrieval
Answers based on approved documents return traceable source citations.
Recovery
The service and approved knowledge base can be restored from the defined backup set.
Model quality
The selected model produces coherent, useful output for the agreed priority workflows.
Ready when you are
Start with one workload.
Tell us who uses it, what it can read, and what must stay inside. We’ll map the build.