FOR OFFICIAL USE ONLY  ·  FOUO  ·  DISTRIBUTION AUTHORIZED TO OPERATIONAL PERSONNEL ONLY
Defence Node // Core
System v3.0 Operational
01 // System Overview

Tactical Intelligence
Workstation

Air-Gapped Local LLM Infrastructure for Defence Operations

A secure, air-gapped-ready intelligence workstation for local analysis of sensitive mission documents using Large Language Models and Retrieval-Augmented Generation — with 100% data sovereignty. Runs on any modern hardware; the application automatically selects the optimal inference model for your system at startup.

Inference Mode
100% Local
Cloud Dependency
Zero
Data Egress
None
Network Required
Offline Ready
02 // System Screenshots Live Build
Hardware-Aware Model Selection — On First Launch Apple Silicon M-series · 48 GB Unified Memory
System Hardware Profile screen showing CPU, RAM, and GPU detection with recommended model gemma4:26b and model tier selection
Note

On first launch, the system scans CPU, RAM, and GPU and recommends the optimal inference model for your hardware. Accept the recommendation or select a different tier — models are downloaded automatically if not already present.

Live Mission Document Analysis — BLUF Extraction Model: gemma4:26b · Op Frost Bite
Defence Node active interface showing a tactical PDF being analysed with BLUF summary, grid coordinates, and key findings extracted
Note

After ingestion the system generates a high-fidelity BLUF summary, extracting grid coordinates, named objectives, risk indicators, and mission discrepancies directly from the source document. All processing is on-device.

03 // Active Capability Matrix
CAP-01

Hardware-Aware Model Selection

Operational

On first launch, the app scans your CPU, RAM, and GPU and recommends the optimal inference tier for your hardware. Accept the recommendation or choose a different model. Downloads are handled automatically if the model is not already present.

CAP-02

Deep Text Intelligence

Operational

Processes large, complex PDF mission documents and delivers high-fidelity BLUF summaries — extracting grid coordinates, risk indicators, and discrepancies. Analysis depth scales automatically with your hardware tier.

CAP-03

Visual Intel Module

Operational

Powered by Moondream. Analyses imagery including satellite maps, drone feeds, and surveillance photos. Performs object and landmark detection, translating pixel data into a clinical, objective text description — fully on-device.

CAP-04

Dual-Stream Memory (RAG)

Operational

Two isolated vector stores: an Ephemeral Mission Archive for the current operation and a Persistent Policy KB for permanent MOD doctrine (JSP 383, Rules of Engagement) loaded via /policy ingestion. The AI automatically cross-references tactical plans against policy to surface legal or ethical constraints.

CAP-05

Integrated OPSEC Layer

Operational

All inference and database communication is bound to loopback 127.0.0.1. Source PDFs and images are overwritten on the physical disk immediately after analysis. Generated SITREPs are securely shredded from the filesystem 60 seconds after creation.

CAP-06

Formal SITREP Export

Operational

One-command generation of sanitized FPDF situation reports from the current session context. Reports are formatted for formal distribution and automatically purged from local storage 60 seconds post-generation.

04 // Inference Model Tiers
Auto-Selected at Startup
Tier Model RAM / VRAM Use Case Example Hardware
High Performance gemma4:26b 40 GB+ Highest fidelity analysis, complex cross-referencing M3 Max 48GB · RTX 4090
Strong gemma3:12b 24 GB+ Solid document analysis and RAG reasoning M2 Pro 24GB · RTX 3090
Moderate llama3.1:8b 16 GB+ Good general performance on standard hardware M1 16GB · RTX 3060 16GB
Entry llama3.2:3b 8 GB+ Basic querying and summarisation on minimal hardware Any modern laptop · 8GB RAM

Visual Intel (Moondream) runs in parallel as a dedicated vision model and is not part of the tier selection above. All models are downloaded via Ollama and stored locally.

05 // Standard Operating Procedures
Unclassified // Releasable
STEP 01
Hardware Profile
On first launch the system scans your CPU, RAM, and GPU and presents the recommended inference model for your hardware. Accept the recommendation or select an alternative tier. The chosen model is pulled from Ollama automatically if not already present before inference begins.
STEP 02
Mission Ingestion
Attach mission-specific PDFs via the file upload control. The system indexes each document into the Ephemeral Mission Archive and immediately generates a BLUF summary with grid coordinates, risks, and discrepancies extracted. Source files are shredded from disk upon completion.
STEP 03
Doctrine Ingestion

Populate the Persistent Policy KB with JSPs, ROE cards, or MOD guidance using the /policy route token. Policy documents persist across sessions and are never auto-purged.

INPUT > /policy [Attach PDF]
STEP 04
Querying & Analysis

Query both archives concurrently with natural language. The model automatically cross-references mission data against policy doctrine to surface compliance deviations, PID requirements, and legal constraints.

QUERY > "What is the PID requirement for technicals at objective ALFA?"
QUERY > "Summarise all sightings of the white pickup truck across the last 3 SITREPs."
STEP 05
Formal Export

Execute the Generate Formal SITREP action to produce a sanitized, standalone FPDF readout from the current session context.

[WARN] PURGE ACTIVE: Generated reports are securely wiped from the local filesystem 60 seconds after creation. Download immediately upon generation.
06 // Security Architecture & Hardware Requirements

OPSEC Protocols

Force Local Binding

All inference and database communication is bound exclusively to loopback 127.0.0.1. External telemetry routes are structurally absent — not disabled, but never written.

Secure Shredding

Original PDFs and images are overwritten on the physical disk immediately after analysis completes. No residual source data persists. Generated SITREPs are wiped 60 seconds after creation.

Local Inference Only

All model inference executes natively via Ollama on the host machine. No tokens, queries, or document content leave the device at any point in the processing pipeline.

Air-Gap Compatible

After initial model download, the workstation operates with no network connectivity required. Designed for deployment in SCIF and air-gapped environments.

Hardware Requirements

Parameter Minimum Recommended
VRAM / RAM 8 GB 40 GB+
Storage 10 GB free 60 GB+ SSD
GPU (NVIDIA) Any 8 GB VRAM RTX 4080+
Apple Silicon M1 8GB M3 Pro/Max 48GB
Quantization IQ4 / 4-bit Q6_K / 8-bit
OS Windows 10+ macOS 14+ / Win 11
07 // Deployment Package Request
Controlled Distribution

Select your target architecture below. Each request routes a deployment bundle inquiry directly to the system administrator and includes pre-configured Ollama manifests, model download scripts, and full setup documentation.

Target Architecture
Windows  /  x86-64
CUDA · DirectML · WSL2 supported
Request — Windows Build
Target Architecture
macOS  /  Apple Silicon
M1 · M2 · M3 · M4 · Unified Memory
Request — macOS Build