Field-ready AI that keeps working when the network goes dark.

In defense and government, teams in the field often work where networks are contested, congested, degraded or cut off entirely. This client, whose name is confidential, needed AI support for English- and French-speaking users operating in exactly those conditions.

Wildstream designed a hybrid edge, fog and core architecture for two deployable AI assistants, one for technical support based on equipment documentation and one for operational knowledge retrieval, built to keep users working through outages of up to 24 hours.

Confidential client

AI Engineering

Up to 24 hours

Edge users keep working through network outages of this length, served by local models and a synchronized RAG cache

Three tiers

Edge, fog and core layers share the AI workload, so each request can run where current conditions suit it best

Four network states

The architecture is built for contested, congested, degraded and disconnected networks, not only for stable connectivity

AI Engineering

Offline Edge AI Assistant

Field users need reliable technical support and knowledge retrieval, in English and French, in environments where connectivity is degraded or completely lost. Their networks can be contested, congested, degraded or fully disconnected, so an assistant that depends on a central server would stop working exactly when users need it most.

Up to 24 hours

Edge users keep working through network outages of this length, served by local models and a synchronized RAG cache.

Two assistants

One supports technical work from equipment documentation, the other retrieves operational knowledge for users at the edge.

English and French

Both assistants serve users in either language, using multilingual models chosen to fit the edge hardware they run on.

What we did

Deployable AI assistants running local multilingual models with a synchronized RAG cache, so users keep working through outages of up to 24 hours. Models are benchmarked for accuracy, latency and power on edge hardware.

Wildstream designed two assistants: one for technical support based on equipment documentation, and one for operational knowledge retrieval. Both run on deployable edge hardware as part of a hybrid edge, fog and core architecture, and every candidate model is also measured for memory use before it is selected. Testing checks how well answers stay grounded in the source material, their overall quality, and whether the local cache stays in sync with the wider system.

How it works

Equipment documentation and operational knowledge form the assistants’ knowledge base
A synchronized copy of that knowledge is cached on the edge device
A user asks a question in English or French
A local multilingual model answers from the cached knowledge, with or without a connection
When the network allows, the cache synchronizes with the wider system

Key capabilities

Synchronized RAG cache

A local copy of the knowledge base stays aligned with the core, so offline answers draw on the same, current content.

Hardware-aware model selection

Multilingual models are benchmarked for accuracy, latency, memory and power use before they are deployed on edge hardware.

Testing under degraded networks

Evaluation reproduces poor or lost connectivity and tracks response quality and service availability across both assistants.

Technology

Edge AI
RAG
Multilingual language models

Impact

Bilingual AI assistants that keep answering in the field, even when the network goes down

Support without a network

Users keep solving equipment issues and finding operational knowledge on site, instead of pausing work until the link returns.

Answers users can trust

Groundedness is measured in testing, so responses stay tied to the equipment documentation and knowledge users already rely on.

Same knowledge everywhere

Synchronization integrity is checked, so people working offline and online consult the same, current version of the content.

AI Engineering

Resilient Edge-Fog AI Architecture

AI workloads in contested networks must cope with shifting bandwidth, latency and data-sensitivity constraints without failing. The system also had to serve two deployable AI assistants for English- and French-speaking users, whose requests cannot simply wait for a stable link to a central server.

Three tiers

Edge, fog and core layers share the AI workload, so each request can run where current conditions suit it best.

Four network states

The architecture is built for contested, congested, degraded and disconnected networks, not only for stable connectivity.

Up to 24 hours

Edge users can keep working through outages this long, because local models and cached knowledge do not need the core.

What we did

A hybrid edge/fog/core architecture that routes each request to the right location based on connectivity, sensitivity and compute, and switches to store-and-forward messaging when the network degrades.

Wildstream designed this architecture behind two deployable AI assistants, one for technical support and one for operational knowledge retrieval. Edge devices run local models and a synchronized knowledge cache for disconnected use, while fog resources take on requests when connectivity, latency and available compute allow. A connectivity manager watches network conditions and changes how messages travel, and testing measures synchronization integrity and service availability as conditions degrade.

How it works

A user sends a request to an AI assistant at the edge
The connectivity manager checks current network conditions
The system weighs latency, data sensitivity, model capability and available compute
The request is handled at the edge or routed to the fog
On a degraded network, messages switch to store-and-forward delivery until the link recovers

Key capabilities

Dynamic workload placement

Each request is routed by connectivity, latency, data sensitivity, model capability and available compute, not by a fixed rule.

Connectivity manager

A dedicated component monitors network conditions continuously and tells the system which communication mode to use at any moment.

Delay-tolerant messaging

When links degrade, real-time messaging gives way to store-and-forward protocols that hold messages and deliver them later.

Technology

Edge computing
Fog computing
Store-and-forward messaging

Impact

An AI architecture that decides, request by request, where work runs, and keeps running when the network does not

Resilience checked in practice

Service availability is tested under degraded conditions, so resilience is measured against the networks users actually face.

Care for sensitive data

Data sensitivity is weighed in every routing decision, helping keep sensitive requests at the tier best suited to handle them.

Data arrives intact

Synchronization integrity is part of testing, so information created offline reaches the rest of the system once links recover.

Better use of compute

Requests use fog compute when it is reachable and stay local when not, making the most of limited deployable hardware.

Room to grow

New assistants or workloads can plug into the same edge, fog and core design without rethinking how connectivity is handled.

Field-ready AI that keeps working when the network goes dark.

Up to 24 hours

Three tiers

Four network states

Offline Edge AI Assistant

Up to 24 hours

Two assistants

English and French

Deployable AI assistants running local multilingual models with a synchronized RAG cache, so users keep working through outages of up to 24 hours. Models are benchmarked for accuracy, latency and power on edge hardware.

Synchronized RAG cache

Hardware-aware model selection

Testing under degraded networks

Bilingual AI assistants that keep answering in the field, even when the network goes down

Support without a network

Answers users can trust

Same knowledge everywhere

Resilient Edge-Fog AI Architecture

Three tiers

Four network states

Up to 24 hours

A hybrid edge/fog/core architecture that routes each request to the right location based on connectivity, sensitivity and compute, and switches to store-and-forward messaging when the network degrades.

Dynamic workload placement

Connectivity manager

Delay-tolerant messaging

An AI architecture that decides, request by request, where work runs, and keeps running when the network does not

Resilience checked in practice

Care for sensitive data

Data arrives intact

Better use of compute

Room to grow

Bring us a challenge like this