Pilot concept · EdgeOps · Offline dual-LLM · NVIDIA or AMD

Two independent AI systems that check each other at the edge

A KingAI pilot concept for higher-trust local AI: two LLMs running on two separate NVIDIA or AMD edge devices. One analyses, the other independently reviews. Agreement raises confidence, disagreement creates a clear review flag, and the full evidence trail is captured locally.

2 Independent systems, no single point of failure
NVIDIA or AMD, vendor-agnostic by design
Offline / air-gapped capable

The concept

A single AI model at the edge has a single point of failure and no second opinion. This concept runs two independent local models on two physical systems, so their outputs can be cross-checked before they become a report, alert or evidence pack.

System A ingests the raw data and produces findings. System B, running a different model on separate hardware, independently reviews the same inputs. Agreement raises confidence; disagreement is surfaced as a decision flag, not buried. Every input, output, confidence score and adjudication result is logged.

Why two, not one

  • A second independent model catches errors a single model would pass through.
  • Two physical devices keep working if one is lost to power, damage or a denied environment.
  • Running on NVIDIA or AMD avoids lock-in to a single supply chain.

Two-system architecture

Two edge devices, two independent models, one adjudication record. Agreement, disagreement, confidence and supporting evidence are captured before the output leaves the local workflow.

SYSTEM A · ANALYST NVIDIA / AMD edge
local model A: loaded (offline)
ingest: sensor, logs, docs, comms
finding: activity of interest · 0.82
finding: logistics movement · 0.68
draft report -> adjudicator
SYSTEM B · REVIEWER separate edge device
local model B: loaded (offline)
re-analyse same inputs, independently
agree: activity of interest
disagree: logistics movement
cross-check -> adjudicator
ADJUDICATION · EVIDENCE CAPTURE

Outcome

Both models agree
High confidence, passes for operator review
Models disagree
Flagged for review with both model outputs retained
One system offline
Other keeps running, degraded but operational
Every step
Logged for audit and after-action review

Why the dual-LLM approach matters

01

Cross-Model Verification

Two independent models must agree before an output is treated as high confidence. Disagreement is a signal, reducing the risk of a single model's hallucination going unchecked.

02

Redundancy & Resilience

Two physical systems mean no single point of failure. If one device is lost to power, damage or a denied environment, the other continues in a degraded-but-operational mode.

03

Vendor-Agnostic & Sovereign

The concept is designed to run on NVIDIA or AMD accelerators, so capability is not locked to a single vendor or supply chain, which supports sovereign resilience.

04

Fully Offline

Both models run locally with no cloud dependency. Sensitive data stays on the devices, suitable for air-gapped, disconnected and controlled environments.

05

Disagreement Detection

Agreements pass forward at high confidence; disagreements and low-confidence findings are captured as review flags with the competing reasoning preserved.

06

Full Audit Trail

Both models' outputs, the cross-check result and the adjudication are logged with sources and timestamps for review, replay and after-action reporting.

Hardware lanes

Near-term, commercial off-the-shelf hardware. The same workflow is designed to run on either vendor, chosen to suit the deployment, budget and supply chain.

NVIDIA lane

Jetson Orin class modules for rugged, low-power edge nodes; RTX workstation or laptop GPUs where heavier local models are needed. CUDA-based local inference, quantised models.

AMD lane

Ryzen AI mini-PCs and Radeon GPUs for local inference, plus AMD-based mini-servers. ROCm / DirectML-style local inference, quantised models.

Mixed / independent by design

The two systems can run different vendors and different model families on purpose, so the review model is genuinely independent of the analysis model.

What a pilot looks like

Pilot shape

  • Two COTS edge devices configured for the target workflow.
  • Customer data, logs, transcripts, sensor feeds, maintenance records or reports.
  • Independent models on separate hardware for a true cross-check.
  • Agreement, disagreement and confidence scoring across both systems.
  • Full evidence capture for audit, replay, handover and after-action review.

Success is measured, not assumed: agreement rate between the two models, how much load reaches a human, time saved versus manual review, and the catch rate on deliberately injected errors.

KingAI can shape this around operational reporting, fleet sustainment, telemetry assessment, exercise capture, firmware review, field notes, local document intelligence or partner-specific EdgeOps workflows. See the broader pilot pathways and EdgeOps for how it fits.

Related KingAI work

EdgeOps

The offline edge-AI reporting and decision-support layer this concept plugs into.

Open EdgeOps

Local & Sovereign AI

Local model setup, private workflows and Australian-owned capability development.

Open Local AI

AI Chip Roadmap

The longer hardware pathway from COTS accelerators toward custom edge devices.

Open roadmap

Defence & AUKUS

How dual-use, resilient, offline AI fits AUKUS Pillar II and Defence-adjacent pathways.

Open Defence

Prove the two-system concept on your problem

Bring one reporting problem, evidence source, log set, transcript set, telemetry stream or document workflow. KingAI can scope a two-device, dual-LLM pilot with cross-checking, disagreement detection, confidence scoring and full evidence capture.