Platform · Private AI Core
Private AI, Right-Sized for the Enterprise.
Run validated 8B–32B models on practical single-GPU or resilient on-premises configurations, sized to context, concurrency, latency and availability.
Design-partner roadmap This build contains no model runtime and calls no model — every model-shaped capability reports an explicit “model not configured” state. The architecture, model classes and hardware figures on this page are intended design and sizing targets: not measured results, and not a supported production specification.
Model, quantization, context length, concurrent agents and response-time target determine the real capacity envelope.
Practical inference infrastructure—not a hyperscale training cluster
- 1Essential
7–9B · 16–24 GB VRAM · 64 GB RAM · 1–2 TB NVMe
Pilot, small team, low concurrency - 2Standard
12–14B · 24–32 GB VRAM · 128 GB RAM · 2 TB NVMe
Recommended normal deployment - 3Professional
24–32B · 48 GB ECC · 128–256 GB RAM · 2–4 TB NVMe
Stronger reasoning, moderate concurrency - 4Resilient
14–32B · two 48 GB GPUs or two nodes · 256 GB RAM · 4 TB+
Redundancy, separation or higher concurrency
Memory planning
Parameters do not equal total GPU memory
Raw weights are only one part of a working inference service, and advertised context limits are not free.
- 01
Weight approximation
FP16/BF16 is about 2 bytes per parameter, INT8 about 1 byte and 4-bit about 0.5 byte.
- 02
Runtime overhead
Quantization metadata, buffers, CUDA graphs, embedding and reranking models, adapters and services consume memory.
- 03
Context and concurrency
KV cache grows with evidence length and simultaneous sessions. Practical limits must be benchmarked.
- 04
14B example
Four-bit weights may occupy roughly 7–8 GB, but that does not make 8 GB VRAM a safe system requirement.
Deployment options
Fit the server to the operating environment
The whole system—airflow, power, drivers, runtime, monitoring, warranty and replacement path—must be qualified.
- 01
Dedicated workstation
A value-oriented option for pilots, small teams and departmental deployments.
- 02
On-premises rack server
A centralized enterprise option for professional GPUs, redundant storage and datacenter controls.
- 03
Existing private infrastructure
GPU virtualization, containers or Kubernetes may be used after passthrough, licensing, driver and runtime qualification.
Availability
Resilience should match the risk
High availability is optional because it materially changes cost and operations.
- 01
Single node
The standard path where queued assistance and planned maintenance are acceptable.
- 02
Warm standby or active/passive
A second approved node shortens recovery without requiring active-active complexity.
- 03
Separated services
Inference and retrieval can be isolated or multiple approved workers placed behind a controlled gateway.
Continuity
Security controls continue when inference is unavailable
The GPU service is not allowed to become a single point of authority for the security platform.
- 01
Deterministic operation continues
Scanners, collectors and policy controls keep running independently of model inference.
- 02
Assistance degrades safely
Requests may queue, use a smaller approved model or require manual review according to policy.
- 03
Recovery is explicit
Queue limits, timeouts, health checks, fallback, recovery objectives and user notices are defined during deployment.
Responsibilities
A shared operating model
Infrastructure ownership does not remove the need for coordinated lifecycle and security decisions.
- 01
SentraMind
Model qualification, crew integration, prompt and retrieval configuration, evaluation, upgrades, rollback packages and product support.
- 02
Customer
Infrastructure, network and identity integration, backup, retention, access approval, power, cooling and license acceptance.
- 03
Shared
Capacity planning, security monitoring, change approval, incident handling and model lifecycle decisions.
Platform · Private AI Core
Size the system around your evidence and workload
The assessment defines the deployment class, model class, tested capacity envelope, availability design and implementation scope before hardware is committed.