Acrodore Tech · AI engineering for regulated organisations

Put AI agents into production
without letting your data leave the building.

We design, build and run AI agent systems inside your own cloud tenant or datacentre. Agents connect to the systems you already run, act under policy you control, and leave an audit trail of every decision they made.

Book a scoping call See engagements

What your security team will ask about

Residency: Runs in your VPC or on-premise. Data does not leave your tenant.
Training: Your data is never used to train models, by us or by the vendor.
Compliance: Built to your controls: Australian Privacy Principles, GDPR, and sector rules we scope with you
Evidence: Immutable, exportable logs of every agent decision
Identity: SSO and RBAC at agent, tool and data level

01What we build

What we build.

Not a chatbot bolted onto a website. A control plane for agents that do real work against your systems, built to be handed over and operated by your own team.

Build

Agent workflows

Multi-step agents with deterministic guardrails, approval gates and clean escalation. Built against your real processes, not a demo script.

Ground

Retrieval and knowledge

Agents answer from your documents, tickets and records, with permissions honoured per user and sources attached to every answer.

Connect

Systems integration

CRM, ERP, ticketing, data warehouse, internal APIs. Agents act in the systems you already pay for, under service accounts you control and can revoke.

Measure

Evaluation and observability

Offline test suites, per-run tracing, and dashboards for resolution rate and cost per outcome. You can see what the agent did and what it cost.

Underneath those sit the parts nobody demos but every security review asks about: orchestration with deterministic guardrails, a simulation engine, end-to-end tracing, hallucination and PII detection, RBAC and SSO, and an immutable audit log. All of it is part of delivery, not a later phase.

02The model layer

An LLM is not a product.
The apparatus around it is.

Most organisations we meet have already wired a chatbot to an API key and watched it stall. It is almost never the model that failed. It is grounding, permissions, cost and evidence, and those are engineering problems with known answers.

Routing

Model choice per step

A frontier model for judgement, a small fast one for classification, open weights for anything that must stay in your tenant. Swapping providers is configuration, so a price change does not reopen procurement.

Grounding

Retrieval that cites

Agents answer from your documents, tickets and records, and every answer carries its sources. An assertion nobody can trace is what kills adoption in month three.

Control

Hallucination and PII guard

Detection and masking on both sides of the call, with policy enforced before a response reaches a user.

Evidence

Tested before it goes live

We simulate each agent across thousands of scenarios, adversarial and prompt-injection cases included, so go-live is a decision your risk function can sign.

Economics

Token cost, attributed

Per-run traces tie latency and spend to the step that caused them. You see cost per resolved ticket, not a single opaque monthly number.

Autonomy

Human in the loop by design

Approval gates, confidence thresholds and escalation that carries full context. Autonomy is a dial you turn up as the numbers earn it.

03Engagements

Four ways in. Start at the one that fits.

Fixed scope and fixed fee at every stage before production, so you can stop after any of them. Figures are indicative bands; your quote follows the scoping call.

Engagement	Duration	Indicative fee	You end up with
Readiness workshop	2 days	Fixed fee	Ranked use cases, data and access gaps, a costed shortlist
Simulated pilot	3–4 weeks	Fixed fee	One agent tested against your historical data, with measured accuracy and deflection before anything runs live
Production deployment	6–10 weeks	Fixed fee	Agents live on a capped scope, integrated, governed, with your team trained to operate them
Managed run	Ongoing	Monthly	We operate, monitor and tune; you get a monthly report on cost per outcome

Third-party cost: Model, hosting and tooling accounts are billed to you directly at cost. We never mark them up, and you see the estimate before we build.
Exit: The configuration is yours. If you take delivery in-house or move to the vendor directly, nothing has to be rebuilt.
Stop points: Each stage ends in a written go or no-go. The pilot exists so a no-go is cheap.

04Deployment

It runs where your policy says it can.

Model A

Managed by us

We host and operate it in our own Australian infrastructure. Fastest to first value where no residency constraint applies.

Model B

Your VPC

Runs inside your own AWS, Azure or GCP account. Your keys, your network boundary, your logs.

Model C

On-premise

Entirely within your datacentre, including open-weights models, where nothing may leave the estate.

AWS

Draw down committed spend

Deployed into your account, the model and infrastructure cost lands against your existing AWS commitment instead of a new supplier line.

Microsoft

Azure, Entra ID, Microsoft 365

Azure deployment with Entra ID single sign-on and integration into Microsoft 365, Dynamics and SharePoint.

Marketplace availability and private-offer terms are set by the cloud provider and depend on your agreement with them. We will tell you at scoping which route is open to you.

05Delivery

First call to first agent in production, in about 90 days.

Week 0

Scoping call

Thirty minutes. What you want to automate, what systems hold the data, who has to approve it. You get a written view of whether this is worth doing, including when it is not.

Week 1–2

Readiness workshop

Two days with your operators and your data owners. Use cases ranked by value against effort, access gaps named, a costed shortlist on paper.

Week 3–6

Simulated pilot

We build one agent and run it against your historical records. You see measured accuracy before a single live transaction is touched.

Week 7–12

Production and handover

Live on a capped scope with human approval on escalations, then your team is trained to operate and extend it. Managed run is optional, not assumed.

06Contact

Book a scoping call.

Thirty minutes, no deck. Tell us what you are trying to automate and we will tell you whether this is worth building, including when the honest answer is that it is not.

Where we work most: banking, insurance, fintech, healthcare, government and private equity. These are the sectors where the governance layer is the reason the project survives its own security review.

07Company

Who you would be contracting.

Acrodore Tech is an independent Australian AI engineering firm. We are not a reseller for a platform vendor and we take no referral fees, so the architecture we recommend is the one we would build for ourselves.

Based: Sydney, New South Wales. We work with clients Australia-wide and remotely.
Independent: No vendor commissions, no marketplace kickbacks, no minimum spend to hit. Tooling is chosen on fit and cost.
Contracting: Fixed-fee statements of work per stage, with a written go or no-go at the end of each. We sign your MSA, NDA and supplier onboarding.
Ownership: Code, prompts, evaluation suites and configuration are delivered into your repository and assigned to you.
Insurance: Professional indemnity and public liability certificates supplied on request during procurement.
Privacy: Australian Privacy Principles apply to everything we hold. See the privacy policy.

08Queries

Before you call.

Are you tied to one vendor?

No. We are independent, and we assemble each system from the tooling that fits your constraints: model providers, vector stores, orchestration and observability. Where an off-the-shelf component is the right answer we use it and tell you what it costs; where it is not, we build it. You get one party accountable for the outcome.

Does our data leave our environment?

In VPC and on-premise deployments, no. The system runs inside your tenant, against your keys, writing logs you hold. Your data is not used to train models in any deployment model.

Which model does it run?

Whichever suits the step and your policy: frontier models from Anthropic, OpenAI or Google, or open-weights models running entirely inside your estate. Routing is configuration, so changing model does not mean rebuilding the workflow.

How do we know it works before it touches customers?

The pilot is deliberately simulated. We run the agent against your historical tickets or records and report measured accuracy, coverage and cost. If the numbers are weak, you stop there having spent a fixed fee.

What happens when an agent gets it wrong?

It escalates with full context rather than guessing. Confidence thresholds, approval gates and capped scope are set during delivery, and every decision is in the audit log, so a post-incident review is reading a record rather than reconstructing one.

Are we locked in to you?

No. The code, prompts and configuration are yours, delivered into your repository. Third-party accounts are held in your name. Handover to your own team is part of the production engagement rather than an upsell, and managed run is offered because some teams want it, not because the build requires us.

What does it cost?

Services are fixed fee per stage, quoted after the scoping call, so there is no open-ended meter running. Model and infrastructure costs are billed to your own accounts at cost, and we estimate them before the build starts.

Put AI agents into productionwithout letting your data leave the building.