AI product development in Australia
ComplxAI is an AI product development company in Australia. We work with founders, SaaS companies and product teams on AI-native applications, AI features inside platforms that already have paying customers, and MVPs designed to become the production system rather than be replaced by it. Design, product and engineering in one senior team, on infrastructure you own.
- Code, prompts and design files are yours from the first paid stage
- Design, product and engineering in one team
- Working software every couple of weeks
Built for the people shipping the product
Users judge an AI product, not a benchmark. It has to answer quickly, say when it does not know, cost less to serve than it earns and keep each customer's data to itself. Those are the problems this work has to solve.
For a system used internally rather than sold, custom AI development is the closer fit; if you are still choosing a use case, start with AI consulting. This page is for a product with customers and a revenue line.
Who we build with
- Founders
- An idea, early customers, a deadline and investors asking what is defensible. Version one must prove demand without boxing in the architecture, and stand up to diligence on the evaluation, the data and the unit economics.
- SaaS companies
- A live platform and a backlog of AI requests that have to fit inside what already works.
- Product teams
- An AI item on the roadmap and nobody who has taken one to production yet.
- Enterprises
- An existing product line and a security team asking about tenancy, residency and audit.
Three shapes of AI product work
The lifecycle below is shared; the entry point differs.
AI-native products
Products that would not exist without the model: a personalised learning tool, a drafting workspace, an assistant that acts inside a workflow. This kind of product is designed around what the model does well and how it fails.
AI features inside existing SaaS
Search that answers, a summary on every record, an agent that closes a ticket, shipped behind your auth and inside your data model. AI feature development is the narrower engagement; SaaS development covers a platform that still needs building.
AI MVPs
A first release scoped to prove one thing to real users, with evaluation, cost telemetry and tenancy handled early enough that it becomes version one of the product. A startup can begin here rather than with a throwaway prototype.
The lifecycle, one decision per stage
In our five-stage engagement, stages 01 to 03 sit inside Discovery and Strategy; from 04 onward the work runs in Implementation blocks.
Define the product: stages 01 to 05
- DiscoveryWhich user, which job, and what a good outcome looks like in their words.
- Product definitionThe one outcome the first release must deliver, and the metric that proves it.
- Technical feasibilityWhether current models can do the job on your data, tested on real examples.
- UXHow uncertainty, latency, citations and corrections appear on screen.
- ArchitectureTenancy, data boundaries, the model interface, orchestration and cost model.
Ship and scale it: stages 06 to 10
- PrototypeOne model, one path, real data, in front of a handful of users.
- MVPThe core loop hardened: golden set, cost telemetry, auth, tenancy, failure path.
- ProductionRouting, caching, monitoring, security review, runbooks and a release process.
- MeasurementQuality, cost, usage and retention read together, by feature and tenant.
- ScalingCheaper models where evals allow, more tenants, new surfaces, the next feature.
An AI SaaS product stack
The stack in seven layers. The model is one row; the product is the other six.
- Product surfacemulti-tenant appReact/Next.js web appMobile appAdmin panelTenant workspacesStreaming UICitationsFeedback capture
- APIcontracts and limitsAPI GatewayCognito authTenant rate limitsUsage meteringWebhooksB2B API
- AI servicesorchestrationStep Functions and LambdaModel routerPrompt registryTool callingGuardrailsApproval stepsResponse cacheFallback paths
- Knowledge and datatenant-scopedRDS/PostgrespgvectorS3 documentsIngestion pipelinePermission filtersConsent and retention flags
- ModelsswappableOpenAIAnthropicGoogleBedrockOpen-weight modelsEmbedding models
- PlatformAWS, in your accountECS/FargateSQSEventBridgeCloudFrontIAM and secretsTerraform/CDKCI/CDRegion you need
- Observability and billingwhat the business seesEnd-to-end tracingEvaluation scoresTokens and cost by tenantUsage analyticsMetering to billingCloudWatch alarms
Lambda and Step Functions carry the bursty work, Fargate the steady services, and Postgres with pgvector handles retrieval until scale says otherwise, all defined as code with CI/CD from the first sprint. The platform is deployed in the region you need; for Australian data residency that is the AWS Sydney region. AI engineering is the discipline underneath the AI services, knowledge and model rows; our AWS consulting practice designs the platform row on its own.
The decisions that separate a demo from a product
Six decisions we work through on every build.
Model and API selection
Chosen per task: OpenAI, Anthropic, Google, Bedrock or an open-weight model, whichever clears your quality bar at the lowest cost and latency. One interface in front of every model, so a swap is a config change and an eval run.
Product UX for AI
Uncertainty shown honestly, streaming so the first words arrive fast, citations users can open, feedback capture that feeds the eval set, and a fallback that finishes the job when the model cannot.
Data: what you have, what you need
An audit of the data the product can use today, the gaps that block the use case, and consent and retention rules, agreed in writing, for what a model may see.
Evaluation and product analytics
A golden set built with your domain experts and a quality score on the product dashboard beside activation and retention, with usage by feature and tenant, latency by stage and retention signals tied to AI interactions, so you see which capability users return for and which they route around. Changes ship when the score holds.
Model costs and unit economics
Cost per request against price per seat, tracked by tenant and feature. Caching, prompt trimming and routing simple tasks to cheaper models make gross margin a design input, not an invoice surprise.
Security and tenancy
Tenant isolation in the data layer, row-level filters on retrieval, secrets kept out of prompts, injection defences on untrusted content, least-privilege IAM and an audit log of every model call.
What an AI MVP has to prove, and what it can defer
An MVP fails in one of two ways: too thin to prove anything, or never shipped. So scoping comes down to two questions: will users pay for the outcome, and can the model deliver it reliably at a cost that leaves a margin? Everything else is deferred by design, not ignored: one model behind an interface that can take another, a tenant ID on every row, every request logged so the eval set builds while you validate demand.
Both columns are legitimate targets. The mistake is building one while promising the other, so we agree the column before the block starts.
| Area | AI MVP | Production-ready product |
|---|---|---|
| Evaluation | Written quality bar and a small golden set, scored before each release. | CI regression suite, judge model checked against human graders, score trend by version. |
| Cost model | Cost per request logged; margin estimated against the intended price. | Routing, caching and tenant budgets; margin tracked by feature and enforced with limits. |
| Tenancy and security | Tenant ID on every row, auth in place, secrets managed, consent rules written down. | Isolation tested, row-level retrieval filters, injection defences, audit log, security review. |
| Monitoring | Request logs and an alert on errors. | End-to-end tracing, token and latency dashboards, quality alarms, runbooks. |
| UX for failure | A clear message and a retry when the model cannot answer. | A fallback that completes the task another way, confidence shown, human escalation. |
| Scaling | One region, one model, provider quota checked. | Queues, async jobs, tenant rate limits, load tested, multi-model fallback. |
Product work with client-reported outcomes
Modedu: a custom SaaS platform for EdTech
A custom SaaS platform and admin panel, AI that generates personalised academic papers for each learner, an agentic-first custom whiteboard, and agentic workflows behind operations. The client reports 2× revenue inside six months, with ARR growing from $400k to $1M.
Read the Modedu case studyLuxpip: a FinTech product suite with a B2B API
For a trading business: a design system, web app, mobile app, admin panel and custom B2B API, with agentic tooling inside internal operations. The client reports a 75% reduction in manual workload and a 40% scale-up.
Read the Luxpip case studyDevelopers who also design the product
The IP is yours from day one
Code, prompts, eval sets, design files and infrastructure definitions sit in repositories and an AWS account you control. No mandatory retainers; take the product in-house at any boundary.
Working software every couple of weeks
Implementation runs in 12-week blocks with fortnightly releases, so real users are on the product while it is being built.
Priced per stage
Two free conversations, then a written price for Discovery, Strategy and each Implementation block. Documentation and handover included.
Design, product and engineering in one team
Naomi on design and product, Eric on engineering, infrastructure and security, Felix on data and machine learning. A small senior team, remote-first across Australia.
Adjacent work we do
AI development
Custom AI systems for internal use rather than for sale.
AI development servicesAI feature development
One AI capability inside the product you already run.
Add AI to an existing productSaaS development
The platform itself: tenancy, billing, admin panel and API.
Custom SaaS platformsAI engineering
The model, retrieval and evaluation layers beneath this stack.
Production AI engineeringQuestions founders and product leads ask
Bring the awkward ones to the intro call. There is no charge for it.
Book a free intro callHow long does it take to get to an AI MVP?
Discovery and Strategy take two to four weeks apiece and produce a scored use-case map and a build plan. The MVP is then the target of the first Implementation block, released to real users fortnightly, with the scope written down before the block begins.
What does AI product development cost?
Each stage carries its own written price, so you never commit to the whole programme at once: two free conversations, a fixed fee for Discovery, an upfront price for Strategy and a price per Implementation block, with an exit at any boundary. For a range before we talk, the AI project cost calculator and the Australian AI development cost report are the places to start.
Do you work with our existing development team?
Yes, and for a live SaaS product we prefer it. Your engineers know the data model and the customers; we bring the AI patterns, evaluation harness and infrastructure, work in your repositories and leave your team able to run the system without us.
Who owns the code, the prompts and the models?
From the first paid stage, you do. Repositories, the prompt registry, eval sets, design files and infrastructure code live in accounts you control. Third-party models stay the provider's, which is why they sit behind a swappable interface.
What happens when a cheaper or better model comes out?
You switch. Every model call goes through one interface with a versioned prompt attached, so a new model gets a run against your golden set, a cost comparison and a staged rollout rather than a rewrite.
Tell us what the product has to do for its users
Bring the roadmap, the prototype or the pitch deck. Two conversations cost nothing, and nothing after them begins until its price is in writing.