Case Study: Bringing locally hosted AI to Defence simulation environments

Client: Department of Defence, Joint Collective Training Branch

Project: Augmented Automation for JLVC

How Profectus developed Augmented Automation, a suite of four locally hosted AI tools that let Defence simulation staff query documents, interrogate logs and alerts, and build JLVC federations in plain English, with every change validated against standards and confirmed by a person before it runs.

Client and context

The Joint Collective Training Branch designs and delivers joint collective training for the Australian Defence Force, using the Defence Training and Experimentation Network (DTEN) and Joint Live Virtual Constructive (JLVC) environments that federate simulation applications across many hosts and domains. Profectus had already modernised this estate, rebuilding simulation software into modular packages, establishing CI/CD automation with Ansible Tower and Git, virtualising legacy platforms and standing up integrated monitoring, which reduced environment deployment from days to under an hour. That work included a prototype, locally hosted LLM orchestration tool that used retrieval augmented generation (RAG) to turn natural language exercise descriptions into build proposals.

The engagement

Profectus was engaged to mature that prototype into a practical, deployable AI capability. The result is Augmented Automation, four modular tools (AA01 to AA04) built on a common, locally hosted LLM foundation. The guiding principle is that automation works best when it makes JLVC’s people more capable, not when it tries to replace them.

The challenge

Bringing AI into a Defence simulation environment placed several compounding demands on design:

  • Offline and sovereign: every tool had to run on JLVC’s own hardware without internet access, with no data leaving the environment
  • Knowledge held in documents: procedures and technical references sat across an in-house repository that was slow to search and relied on individual expertise
  • Specialist-only visibility: logs, metrics and alerts across hosts and applications needed specialist skills to interrogate
  • Complex federation builds: each JLVC host must carry a correct hostname, IP address, MAC address, VLAN and resourcing to application minimums, so manual entry risked clashes, errors and rework
  • Trust and control: AI could not be allowed to make unreviewed changes to live systems

Our approach

Profectus built each tool as a self-contained, offline-deployable application, tested incrementally and designed to work together across the platform.

AA01: a private chat interface for local LLMs

A web chat for Ollama-hosted models, packaged as a single Docker container that moves to offline hosts by tarball. A model selector lets teams run small models on modest hardware or larger models where the hardware allows.

AA02: answers from JLVC’s own documents

Indexes the in-house document repository and answers with its sources cited under every reply, saying so when nothing relevant is found rather than guessing. New and changed documents are re-indexed automatically, staff can save important facts as a shared team memory, and an encrypted vault holds secrets away from chat.

AA03: logs and alerts in plain English

Queries error-level logs in Loki and alert history from Prometheus and Alertmanager, answering only from JLVC’s own logs and metrics. Staff can filter by host or time window, chart CPU, memory, disk and network against thresholds, and track individual JLVC applications as well as hosts. The full stack is packaged for offline sites.

AA04: run and manage JLVC builds from chat

Connects to Jenkins to list, run and monitor pipelines (build, install, uninstall and security check), with plain-English explanations of failed builds. Hosts are added singly, by CSV or Excel upload, or as predefined groups, with hostnames, IPs, MACs and VLANs filled in automatically and every value checked against an editable JLVC rules file and existing systems before any change is committed to the version-controlled build repository. AA04 also passes document questions to AA02 and alert questions to AA03.

What we delivered

  • A working, tested suite: four integrated tools developed and tested on JLVC infrastructure, ready for production deployment
  • Responsible AI by design: no LLM in any write path, allowlisted jobs and files, single-use confirmation tied to the user, previews before every commit and a separate audit log
  • Standards enforced automatically: build parameters held in one editable rules file, so every build adheres to the JLVC standard and the standard can change without code changes
  • Sovereign and portable: every component runs offline on Defence hardware, with jobs, pipelines and agents able to transfer to other offline sites
  • Modular growth path: new capability is added by extending pipeline tools, with automated application configuration to standards and per-application tracking and execution identified as next steps

Systems and methods

Locally hosted LLM (Ollama) · RAG · Docker · FastAPI · Loki, Prometheus, Alertmanager and Alloy · Jenkins · Gitea · VMware vSphere · AWX

The result

Augmented Automation gives the Joint Collective Training Branch a sovereign, offline AI capability that is developed, tested and ready for production. It extends Profectus’s earlier modernisation of DTEN by putting documents, system health and federation builds within reach of plain-English requests, while keeping people firmly in control of every change. Once deployed, it is designed to reduce reliance on specialist knowledge, cut build errors and speed the preparation of JLVC federations. Operational results will be reported following production deployment.