Platform · Private AI Core
The Right Model, Grounded in Your Approved Knowledge.
A model-independent design selects the smallest validated model that meets language, security quality, performance and licensing requirements.
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.
Current approved knowledge is retrieved with provenance; private documents are not automatically trained into model weights.
Versioned, permission-preserving model service
- Approved base model
Commercial rights, provenance, local runtime and security-task quality are reviewed.
- Governed private knowledge
Policies, procedures, asset context and permitted security references are retrieved with source permissions.
- Task prompting
Versioned instructions and output schemas constrain each specialist workflow.
- Optional validated adapter
LoRA or QLoRA is introduced only when a controlled evaluation proves measurable benefit.
- Signed release
The exact model stack is versioned, approved, monitored and rollback-capable.
Model policy
Open-weight, commercially deployable, independently validated
“Open source” has a precise and disputed meaning for AI. SentraMind uses checkpoint-specific commercial and technical review.
- 01
Representative ecosystems
Selected Qwen, Mistral and IBM Granite releases may be candidates; other ecosystems are considered when their exact terms and artifacts pass review.
- 02
License is checkpoint-specific
Base model, fine-tune, adapter, quantization and redistribution terms are reviewed together with customer legal acceptance.
- 03
No vendor lock-in claim
A provider-neutral gateway enables evaluated upgrades, replacements and rollbacks without treating every model as interchangeable.
Parameter tiers
Model size is a design choice, not a score
Larger models can improve difficult reasoning but consume more memory, power and response time.
- 01
Efficient · 7–9B
Classification, extraction, summarization and first-pass triage.
- 02
Balanced · 12–14B
Recommended starting tier for multilingual security reasoning and agent assistance.
- 03
Enhanced · 24B
Complex reports, code reasoning and stronger instruction following.
- 04
Advanced · 30–32B
Difficult correlation and deeper analysis where latency permits.
- 05
Large-scale · 70B+
Specialized option, generally outside the affordable default design.
Selection gate
Supported means tested for the intended deployment
Marketing names are insufficient. Candidate configurations must satisfy a versioned acceptance gate.
- 01
Grounding quality
Finding accuracy, evidence citation, unsupported-claim rate and structured-output validity are measured.
- 02
Security behavior
Prompt injection, knowledge poisoning, sensitive-data handling, refusal and authority boundaries are tested.
- 03
Operational fit
First-token delay, completion latency, peak VRAM, concurrency, restart and recovery are measured on the exact stack.
- 04
Language and task fit
Required languages and real SentraMind tasks—not generic leaderboards—determine acceptance.
Knowledge strategy
RAG first; fine-tuning only when proven
Retrieval keeps knowledge current, permission-aware and easier to remove or update than information embedded in weights.
- 01
Approved sources
Customer policies, asset context, incident procedures, internal documentation and permitted OWASP or MITRE material.
- 02
Permission-preserving retrieval
Collections are segregated and results inherit source authorization and provenance.
- 03
Controlled adaptation
A defined dataset, evaluation and rollback plan are required before any task adapter is accepted.
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.