Architecture

Two planes, one engineering discipline

Scrutari runs on two data planes built to the same standard. A hermetic Rust AI gateway terminating post-quantum TLS in front of your AI and API traffic, and offline-first inference on ruggedized hardware where connectivity is not guaranteed. Both are engineered so failure is contained, not catastrophic.

AI Gateway · Data Plane

A hermetic Rust data plane in front of your AI traffic.

The gateway terminates hybrid post-quantum TLS on the public edge, runs every request through a governance layer (token quotas, PHI redaction, signed audit), then fans out to OpenAI, Anthropic, Gemini, and Bedrock behind one OpenAI-compatible API. One self-contained process: no garbage collector, no interpreter, a single attack surface.

Hybrid PQ-TLS termination

X25519MLKEM768 terminated at the public edge. TLS 1.3 only, no classical-only fallback. The upstream hop re-encrypts over classical TLS 1.3 to your origin.

Fail-closed governance

A request that cannot reach its quota store, key store, or PHI scanner is refused, not forwarded. For a compliance boundary, blocking is the safe failure mode, never leaking.

Signed, tamper-evident audit

Every AI request writes an append-only, Ed25519-signed row, Merkle-anchored on Growth and Enterprise. Retention you control, queryable from your own SIEM.

Multi-provider routing

OpenAI, Anthropic, Google Gemini, and AWS Bedrock behind one OpenAI-compatible, Anthropic-native API. Bring your own provider keys, encrypted in a key-management boundary.

Per-tenant isolation

Tenant-scoped row-level security isolates multi-tenant traffic at the database layer. Per-tenant monthly token quotas enforce on calendar-month UTC windows.

Single-process Rust runtime

The data plane is one self-contained Rust process: no garbage collector, no GC pauses, a single attack surface. Memory-safe by construction.

Command Center

Actionable Safety Telemetry

The system doesn’t just detect defects. It creates an immutable, verifiable trail of safety events. The Command Center provides site managers with immediate, centralized visibility across all deployed edge nodes.

Live Anomaly Log

Automatically tracks and timestamps every detected defect, translating visual data into actionable maintenance tickets in real time.

Fleet Management Ready

Designed to scale via Over-The-Air (OTA) updates to hundreds of connected mining sites globally. Zero-touch deployment.

False Positive Elimination

Engineered to cleanly differentiate between human workers and infrastructure defects. Tuned for precision, alerts are actionable, not noisy.

COMMAND CENTER
Demo
LIVE
14:32:07Class 4 corrosion detectedGantry B-7
14:31:42Fastener vibration alertCrane A-3
14:30:15Bolt elongation flagTower C-12
14:28:53Spalling detectedColumn D-4
14:27:11System health checkAll nodes
FLEET STATUS
47
ONLINE
3
ALERTS
2
OFFLINE
Simulated data • Real product interface • Request live demo

See the Command Center Demo

Watch our dark-mode Command Center actively tracking infrastructure defects in real time. Timestamped anomaly logging, detection overlays, and confidence scoring.

Request Demo Access

What Rust Gives Us

These are properties of the language, not marketing claims.

Memory Safety at Compile Time

Buffer overflows and use-after-free eliminated before code runs. The compiler enforces safety, no runtime checks, no garbage collector.

Deterministic Performance

No garbage collector means no surprise pauses during inference. When a falling object alert has milliseconds to fire, you need deterministic latency.

Compiles to ARM / Edge

Same codebase compiles natively for edge ARM hardware. No Python interpreter, no JVM, no container overhead on the edge device.

Thread Safety by Default

Multiple camera feeds processed concurrently with compile-time guarantees. Data races aren’t unlikely, they’re structurally impossible.

Strong Type System

Sensor data, alert payloads, and equipment IDs are strictly typed. Mismatches caught at compile time, not in production at 3 AM.

24/7 Uptime

No memory leaks means no degradation over weeks of continuous operation. The inference engine runs as long as the hardware does.

Edge Deployment Model

Each monitoring point consists of industrial-grade cameras feeding a ruggedized AI compute module. All inference runs locally. The only outbound traffic is kilobyte-sized encrypted alert payloads.

Fleet management is handled via secure OTA updates. Model improvements, configuration changes, and security patches deploy to the entire edge fleet simultaneously without physical site access.

Ruggedized AI Hardware

Enterprise-grade compute module rated for extreme temperature ranges and heavy vibration environments.

Secure OTA Updates

Signed, encrypted firmware updates. Zero-touch deployment with automatic rollback on failure.

Local-First Storage

Embedded database with offline-first operation. Secure bidirectional sync when connectivity returns.

DEPLOYMENT ARCHITECTURE
[Camera Array]────▶Edge AI Module
LOCAL INFERENCE
Real-time • Proprietary GPU backend
Encrypted Alert── KB ──▶Command Center
Work Order────▶[SAP PM / Maximo]
No video streaming • KB payloads • Cryptographic audit trail

Due Diligence Ready

Designed to pass Tier 1 mining company procurement and vendor qualification processes.

SCRUTARI AI LLC, US-domiciled, strict FCPA compliance
AES-256-GCM authenticated encryption for all transmitted payloads
Hybrid post-quantum key exchange in transit: X25519 + ML-KEM-768, NIST FIPS 203
Signed OTA updates with automatic rollback on verification failure
Cryptographic audit trail, hash chain for every safety event
No raw video data leaves the site, only KB-sized alert metadata
Hardware-agnostic installation via Tier-1 electrical contractors
ESG-ready reporting with export formats for regulatory submissions
Role-based access control for multi-site fleet management

Request the Full Architecture Whitepaper

Our NDA-protected engineering specifications include detailed deployment architecture, hardware configurations, inference benchmarks, and API integration documentation.

Request Under NDA