Skip to main content

Architecture

How the system works

From raw sensor data to structured operational decisions — the complete technical architecture of the KrinisAI intelligence stack.

Stack

Four-layer architecture

VisionOS

Perception Layer

Industrial cameras and IoT sensors feed raw data into on-premise GPU nodes. VisionOS runs inference at line speed with sub-50ms latency. No cloud — data stays on-site.

Industrial IP camerasEdge GPU nodes (NVIDIA)Custom CNN/ViT modelsFrame buffer and preprocessing pipeline
Shared

Signal Fusion Layer

Visual outputs from VisionOS are combined with SCADA, PLC, MES, and manual signals. Each source is assigned a trust weight based on its reliability and recency.

SCADA/PLC integrationMES connectorsIoT sensor streamsTrust-weight configuration engine
BrainOS

Reasoning Layer

BrainOS ingests the fused signal stream and applies authority-bounded generative reasoning. Decisions are scoped to explicitly configured limits — no open-ended autonomy.

Authority configuration engineGenerative reasoning modelUncertainty quantificationDecision trace logger
BrainOS

Action Layer

BrainOS emits structured actions — line halts, supervisor briefings, maintenance work orders — or escalates to a human when confidence is insufficient.

Action executorEscalation engineOperator dashboardAudit log export

Security Architecture

Safety by design

Security and safety are not features added at the end of development. They are structural constraints baked into the architecture from day one.

No cloud dependency

All inference and reasoning occurs on-premise. Sensitive production data never leaves the facility network. Air-gapped deployments are fully supported.

Authority limits are immutable at runtime

Authority bounds are set during deployment configuration and cannot be modified at runtime by the system itself. Changes require operator authentication.

Cryptographically signed audit logs

Every decision emits an immutable, timestamped log entry. Logs are tamper-evident and exportable for compliance and incident investigation.

Principle of least privilege

Each system component operates with the minimum permissions required. VisionOS cannot write to BrainOS decision logs. BrainOS cannot access raw camera feeds directly.

Fail-open for safety, fail-closed for authority

In degraded states, safety-critical functions default to safe states. Authority claims are rejected rather than relaxed when the system cannot verify them.

OWASP-aligned web interfaces

All operator dashboards and API surfaces are built to OWASP ASVS standards — including CSP, input validation, rate limiting, and secure session handling.

Talk to the engineering team

Schedule a technical deep-dive to understand how the architecture fits your operational environment.