Agentic Services & Emerging Technology
AI and ML as delivery tools and embedded in solutions. DOGRAG methodology for grounded AI outputs.
Overview
We were building on AWS when it was S3 and EC2 and not much else. We used blockchain early enough to find out, at our own expense, which problems it actually solves. We had machine learning in production government systems well before the current wave.
The pattern is adopt early, then be honest about what you found. That is how we are treating agentic services.
What We Offer
AI presents two opportunities for our clients.
As a delivery tool. Our developers use AI coding assistants daily (Claude Code, Aider, Cursor, Cline). The same coding standards, testing, and review practices apply to AI-assisted code as to hand-written code. AI-assisted development is systematised across the organisation with consistent quality expectations, not left to individual experimentation.
Embedded in delivered solutions. We build AI capabilities into client systems where they deliver measurable results:
| Application | Example | Evidence |
|---|---|---|
| Predictive compliance | ML models targeting audit resources at highest-risk entities | DAFF Export Assurance |
| Automated classification | ML-driven auto-classification against regulatory taxonomies | Finance RMaaS, 500M+ records |
| Process mapping | Generative AI for business process analysis | ABF Trade Modernisation |
| Document processing | AI-accelerated compliance assessment and entity resolution | RBA Online, 600+ member companies |
| Conversational interfaces | LLM-powered access to complex data for non-technical users | DoH Ask MCF |
We deliver on Microsoft Azure AI Foundry and AWS AI/ML services, selecting the platform that fits the client’s existing cloud ecosystem.
Bring the model to the data
Plenty of government clients cannot send sensitive documents to a cloud AI service, and no amount of contractual comfort changes that. So we put the model inside their boundary instead: open-weight LLMs in the 70B class, running on hardware they control. System inventories, capability maps, strategy papers and architecture documentation stay on the network.
There is a design trick that makes this work better than it sounds. Rather than handing a whole problem to one large model and hoping, break it into many small focused extractions that build a knowledge graph piece by piece, then run ordinary deterministic queries over the graph.
The results are more reliable than single-pass processing by a frontier model, and you can audit how each one was reached.
DOGRAG
Standard RAG retrieves relevant context and feeds it to an LLM. Our Domain Ontology Grounded RAG (DOGRAG) methodology adds an ontological grounding layer that constrains AI outputs to established domain frameworks. This prevents the common failure mode where LLMs produce plausible but architecturally inconsistent outputs.
First the ontology gets defined, and which one depends on the work. Requirements analysis uses epics, features, stories and personas. Architecture uses C4. Decomposition uses bounded contexts, aggregates and services.
Retrieval is then filtered and structured to that ontology, so the model gets the right framework and not merely the right documents.
Generation is constrained against it, and anything that drifts outside is rejected and regenerated rather than quietly accepted.
Finally each generated component is checked against its own scope, interfaces and responsibilities. A component can be individually plausible and still wrong for where it sits.
Principles
- Guardrails before acceleration. Bounded contexts, interface contracts and test coverage come first. AI amplifies whatever practices it finds, and it does not care which ones.
- Developer ownership. Whoever commits it answers for it, however it was written. If the team cannot understand, review, test and maintain a generated component, it does not go in.
- Honest about the frontier. We will tell you what works now, what is still experimental, and where reliable AI-assisted delivery currently runs out.
Policy Alignment
The Policy for the Responsible Use of AI in Government sets the expectations on human oversight, transparency, accountability and impact assessment, and the AI Technical Standard adds 42 statements across the system lifecycle. We build to both.
Nothing about AI-assisted work sits outside the Essential Eight or the ISM. It runs in the same security framework as everything else we do.
Evidence
Export Assurance put machine learning behind predictive compliance across the $80 billion agricultural export market. Records Management as a Service used it for auto-classification over 500 million records. ABF trade modernisation used generative AI for process mapping and the future-state cargo architecture. The RBA Online platform is current work: entity resolution, document assessment and compliance automation across more than 600 member companies.
Michael Nelson is a published author on building with LLM agents. Luke Tankey established the AI-assisted coding practice at DEWR, including the guidelines and tooling. Niko Kresic has built LLM proofs of concept on Langchain.
Day to day the coding assistants are Claude Code, Aider, Cursor, Cline and Zed. Frameworks are Langchain, RAG pipelines and our own DOGRAG implementation. Cloud AI runs on Azure AI Foundry and Copilot Studio, or AWS SageMaker, Comprehend and Rekognition. Locally we run Llama 3.3 70B, Qwen 2.5 72B and DeepSeek-R1 70B, with constrained decoding through Outlines or llama.cpp grammars, served by vLLM or Ollama.