AI Agent Development Company in Canberra

Fix the classification before anything else, because it drives every other line item: a scoped agent proof of concept inside an OFFICIAL: Sensitive environment over eight to twelve weeks, covering PSPF risk assessment, Essential Eight gap analysis, an Algorithm Charter conformance plan, and a hosted demo inside Azure Australia Central 1, or a production agent rollout for a department (Services Australia case-officer copilot, ATO document-summarisation agent, DFAT cable-drafting agent) including IRAP PROTECTED assessment support, AGSVA-cleared delivery, and an AI Ethics Framework self-assessment. SECRET-adjacent agent work inside agency-owned GPU enclaves runs higher because of physical access, cleared-personnel cost, and longer assessor cycles. All proposals are fixed-fee in AUD with no T and M creep, billed against the Digital Marketplace BuyICT panel where possible.

Codazz is structured as a Brisbane-based account team plus Chandigarh build hub. AGSVA Baseline, Negative Vetting 1, and Negative Vetting 2 cleared engineers are sourced through partner panels in Canberra and Brisbane, with project-specific clearance sponsorship handled by the prime contractor or the agency directly. We are honest about AGSVA backlogs: NV1 routinely takes six to nine months and NV2 longer, so we structure projects so that cleared work is bounded and non-cleared engineers handle infrastructure, observability, and non-sensitive prompt engineering. Defence Industry Security Programme (DISP) membership is sponsored through partner primes when scope requires it. We do not over-claim clearance status — the Defence Security Vetting Service investigates every assertion, and so should you.

The Australian Government's Voluntary AI Ethics Framework (released 2019, updated through 2024) sets eight principles: human-centred values, fairness, privacy and security, reliability and safety, transparency, contestability, accountability, and net benefit. Each agent build ships with a conformance matrix mapped to those principles plus the new mandatory AI assurance framework signed off by Digital Transformation Agency guidance in 2024. Human-in-the-loop controls, override pathways, decision logs, and notification to affected parties are baked in — these are non-negotiable after the Robodebt Royal Commission. We also map against New Zealand's Algorithm Charter for Aotearoa when a project has trans-Tasman exposure (Five Eyes intelligence sharing, ANZ defence cooperation, dual-sovereign procurement).

Yes. The Hosting Certification Framework (HCF) certifies cloud providers and data centres as Certified Strategic, Certified Assured, or Certified Sovereign for federal workloads. Azure Australia Central 1 and 2 hold Certified Strategic status for IRAP PROTECTED, AWS Sydney ap-southeast-2 is Certified Assured for OFFICIAL: Sensitive and below, and a small set of Canberra-region data centres are Certified Sovereign for the most sensitive workloads. We structure every Canberra agent deployment against the HCF tier required by the data classification, with no exceptions and no implicit downgrades. Where the data is SECRET, we deploy on agency-owned hardware inside ASD-accredited spaces and use frontier models only as offline-distilled smaller models inside the enclave.

Federal agent timelines are dominated by procurement and assessor cycles, not engineering. A BuyICT-panel SOW typically takes four to eight weeks to award. PSPF and Essential Eight pre-assessments take three to five weeks. IRAP PROTECTED assessment for a new system takes ten to sixteen weeks with the queue. Build itself, for an OFFICIAL: Sensitive case-officer agent, is twelve to twenty weeks of engineering wrapped in evidence collection. Realistic end-to-end is nine to fourteen months from RFP to production for OFFICIAL: Sensitive and twelve to twenty-four months for PROTECTED. We deliver fixed-fee against these realities — anyone promising six-week production for a federal department is selling a demo, not an agent.

When a project requires genuinely novel agent reasoning (multi-step planning under uncertainty, RL-on-document workflows, formal verification of agent actions for safety-critical decisions), we collaborate with ANU's School of Computing, the ANU Tech Policy Design Centre, or UNSW Canberra at ADFA cyber-defence research groups. ANU is G8 and consistently ranked in the global top fifty for AI research. Most federal agent work is engineering, not research, and we will say so up front rather than over-scoping academic collaboration. Where collaboration genuinely fits, we structure it through formal research agreements with IP, publication, and citizen-eligibility terms that survive AGSVA and Defence review.

Every agent that touches a decision affecting a member of the public ships with a documented human-in-the-loop override, a complete decision log, a notification pathway to the affected person, and a contestability surface mapped to existing review and appeal rights (AAT, internal reconsideration, ombudsman pathways). We do not build agents that issue debt notices, eligibility denials, or compliance actions without a human officer making the binding decision. The Royal Commission into the Robodebt Scheme made this non-negotiable for Australian federal AI, and the agency's general counsel reviews the human-in-the-loop matrix before go-live. Anything that smells like automated adverse action gets stopped at design review, regardless of efficiency gains promised.

Yes, through panel arrangements. Codazz partners with primes already on BuyICT (Digital Marketplace), the Department of Finance ICT Services Panel, Defence panels including ICT Provisioning Arrangements (ICTPA), and the Services Australia ICT Services Panel. We do not currently hold direct BuyICT entitlement as a small business in our own name, and we will be transparent about that in any procurement conversation. For agency clients who prefer direct contracting under threshold (typically AUD $80,000 limited procurement), we contract straight in AUD with fixed-fee scope, an Algorithm Charter conformance matrix, and an Essential Eight self-assessment included.