CLASSIFICATION POSTURE · PROTECTED
DEFENCE · AI SYSTEMS
AI software for the workloads you cannot afford to get wrong.
We build production AI applications for defence customers, not pilots, not demos. Retrieval-augmented generation, semantic search, document understanding, and decision-support systems engineered for auditability, attribution, and human-in-the-loop review. Currently in operation inside an air-gapped Australian Armed Forces environment.
An AI system in a defence environment is not measured by its benchmark. It is measured by what happens when it is wrong, and by whether the operator can tell.
PRACTICE OVERVIEW
Production AI, designed for the constraints defence workloads actually have.
Most generative AI failures in defence look the same. The model produces a plausible answer the operator cannot verify. The system pulls from a corpus the customer cannot audit. The pipeline calls out to a hosted model it should not have. The interface treats the answer as the product, when the answer is only the start of a decision the operator is responsible for. RankSaga's AI practice is built around closing each of those failure modes, not as policy, but as engineering.
We build AI applications that operators can use for real work because the system is designed around three obligations: every output is attributable to a source, every action a human takes against the output is logged, and every model in the pipeline is one the customer is allowed to run on the data the system handles. None of those are negotiable. They are the ground floor of every system we ship into a sensitive environment.
The technical surface is broad, retrieval-augmented generation over authoritative document corpora, semantic search across heterogeneous mission data, document understanding for classified and unclassified materials, decision-support systems that summarise, score, and recommend, and operator-facing agent consoles that combine all of the above. The common thread is that the systems are built and deployed inside the customer environment, with the constraints of that environment shaping the architecture from day one.
We are equally pragmatic about model choice. We deploy sovereign-hosted foundation models when the customer's residency posture requires it, customer-fine-tuned models when domain accuracy demands it, and open-weight models we run inside the customer environment when the workload is too sensitive for any external inference path. Model selection is an engineering decision driven by the workload, not a vendor relationship.
WHAT WE BUILD
AI applications, by the work they do.
01 / Capability
Retrieval-Augmented Generation
RAG systems over authoritative customer corpora, policy libraries, doctrine, mission data, briefing materials. Source-attributed answers, hybrid retrieval, customer-controlled vector stores, and an evaluation harness the customer can run themselves.
02 / Capability
Semantic Search & Knowledge
Search across heterogeneous mission data with embedding-based retrieval, knowledge-graph traversal, and hybrid reranking. Indexes the customer owns; relevance the customer can audit.
03 / Capability
Document Understanding
Extraction, summarisation, classification, and entity resolution across structured and unstructured documents, including multi-page, multi-format, classified-handling material, at scale and with auditable lineage.
04 / Capability
Decision-Support Models
Scoring, prioritisation, and recommendation systems that surface candidate decisions with the evidence behind them. Designed to inform an operator, not replace one.
05 / Capability
Agent Consoles & Operator UI
The interface is the product. We build operator-facing applications, multi-step agent consoles, briefing tools, after-action review surfaces, that present AI output with the density, latency, and verifiability operational use demands.
06 / Capability
Evaluation, Red-Teaming, and Guardrails
Customer-runnable evaluation harnesses, red-team test suites, and policy enforcement layers that constrain what the system can do, what it can say, and what it must surface to the operator.
OPERATING MODEL
Built in the customer environment, hardened from day one.
Every defence AI system we ship is built inside the customer's production environment, against the customer's data, with the customer's operators in the loop from week one.
01 / Step
Workload & Threat Mapping
We start by mapping the workload, what the operator is actually trying to do, what data is available, what classification it carries, what failure modes are tolerable, and what controls the environment enforces. The model and architecture choices fall out of that, not the other way around.
02 / Step
Build Inside the Environment
We provision into the production target, ingest authoritative customer data under the correct controls, and ship working AI software in operator hands within weeks. Auditability, attribution, and human-in-the-loop review are designed in from the first sprint.
03 / Step
Operate, Evaluate, Iterate
We stay deployed. We run the evaluation harness against new releases, ingest field feedback, monitor for regression and drift, and harden the system against the threats it actually sees. The team that built the system is the team that operates it.
WHAT YOU GET
Software, evidence, and the team that built it.
01 / Deliverable
Working AI Application in Production
Operator-facing AI software running in the customer environment against real data, not a notebook, not a demo, not a hosted SaaS surface.
02 / Deliverable
Evaluation Harness
A test harness the customer can run themselves to evaluate new model versions, new prompts, new retrieval configurations. The customer is the judge of quality, not us.
03 / Deliverable
Audit & Attribution Surface
Every output traceable to its sources. Every operator action against the output logged. The audit surface is part of the system, not an external add-on.
04 / Deliverable
Operations Runbook & Embedded Engineers
A documented operations posture and the engineers who can hold it. Patching, model updates, drift monitoring, and incident response, by the people who built the system.
REFERENCE
Live for the Australian Armed Forces. Air-gapped.
The AI application we built and operate for the Australian Armed Forces runs inside an air-gapped environment, on Microsoft Azure sovereign infrastructure, with model artefacts and inference paths designed for an environment with no internet route.
- ·Source-attributed retrieval over authoritative customer materials.
- ·Operator console with human-in-the-loop review at every consequential step.
- ·Customer-controlled model lifecycle, including offline updates inside the enclave.
- ·Continuous evaluation harness operated by the customer.
RELATED CAPABILITIES
Where defence AI meets the rest of the stack.
Adjacent
Mission Software Engineering →
When the AI surface is part of a broader operator-facing application, interfaces, integration layers, decision-support backends.
Adjacent
Air-Gapped Deployment →
When the AI system has to operate inside a disconnected enclave with no inference path to the outside world.
Adjacent
Microsoft Azure (Sovereign) →
When the deployment target is sovereign Azure with IRAP-aligned controls and PROTECTED-class workload handling.
QUESTIONS
What customers ask first.
Can the AI system run without an internet connection?+
Yes. We have shipped AI systems into air-gapped environments for the Australian Armed Forces. Model artefacts, vector indices, and inference paths run inside the enclave, with offline update flows and a hardened supply chain for model and dependency updates.
Whose models do you use?+
Workload-driven choice. We have shipped systems built on sovereign-hosted foundation models, on customer-fine-tuned variants, and on open-weight models we deploy and harden inside the customer environment. The decision is made on residency, auditability, and operational fit, not on vendor preference.
How do you handle hallucination and incorrect output?+
Two layers. First, retrieval-augmented architectures so every output is anchored to a source the operator can inspect. Second, an operator interface that surfaces the source, the confidence, and the alternative candidates, designed so the operator is the decision-maker, not the AI.
Do you provide an evaluation framework, or do we have to build one?+
We deliver a customer-runnable evaluation harness as part of the engagement. Quality is the customer's call, they need to be able to evaluate new model versions, new prompts, new retrieval configurations against their own benchmarks without us in the loop.
Can you integrate with our existing data infrastructure?+
Yes. Our engagements include the integration layer, connecting to systems of record, document repositories, message buses, and identity providers. The AI surface is one part of a larger application; we build the rest of it as well.
ENGAGE
If the workload matters, the AI behind it should be built like it does.
We are most useful when the customer cannot afford a hallucination, when the data cannot leave the environment, and when the people responsible for the decision need to be able to see why the system said what it said.
ENGAGE
Bring us in on the problem before it has a name.
We work best when we are embedded early, alongside the team that owns the mission, the data, and the operational risk. Government, commercial enterprise, or defence: if your environment is regulated, sensitive, or air-gapped, that is where we are most useful.