Custom AI development in Australia
ComplxAI is an AI development company in Australia. We build custom AI applications end to end: the interface people use, the API and orchestration behind it, the knowledge pipeline it draws on and the AWS platform it runs on. You own the IP from day one of any paid stage.
- Working software every couple of weeks
- Your code and your IP from the first paid stage
- Designed, built and operated by the same senior team
AI development that ends in a running system
AI development is, in practice, ordinary software development with a model inside it. The model is the small part; data plumbing, orchestration, permissions, tests and deployment decide whether the system still works a year on.
This page is about building the whole system, greenfield or inside software you already run. Still choosing which use case deserves the budget? Start with AI consulting in Australia. Have a live system that needs hardening? See production AI engineering.
Custom AI applications, products and internal tools
Six shapes the work takes; a brief often spans several.
Custom AI applications and internal tools
Complete applications with a model at the core: front end, authenticated API, background workers, database, as infrastructure-as-code from the first sprint. Or the same capability inside the software your team already opens, inbox, CRM, document store, admin panel, measured on hours handed back and wired in through our AI integration services.
LLM-powered systems
A language model reading, writing, extracting or deciding at volume: versioned prompts, schema-forced outputs, tools for facts and actions. Our generative AI consulting settles which tasks suit one.
AI agents
Agents that plan, call tools and act inside your systems, pausing for a person when an action has consequences. Every step traced; writes reviewed first or reversible by design. See AI agent development.
RAG pipelines and knowledge assistants
Retrieval-augmented generation over documents, tickets, contracts and databases: ingestion, chunking, embeddings, hybrid search, reranking and permission filters so people only see what they may. On top, search that answers in sentences and cites its passage, with pgvector in Postgres or a managed vector database at scale, freshness jobs and an escalation path when the answer is missing.
AI SaaS products
Multi-tenant products where the AI is the product: billing, tenant isolation, usage metering, model-cost telemetry per customer, admin panel. Positioning and roadmap sit with AI product development.
AI features inside existing products
One capability added to a product you already ship, a summariser in the inbox or an extractor in the upload flow, without disturbing release cadence. See adding AI features to an existing product.
The engineering underneath every build
Sized to the brief and priced into the plan, not added after the demo.
- Data pipelines
- Ingestion from S3, email, document stores and internal systems; cleaning, chunking, metadata; refresh jobs so the model reads current data.
- Backend architecture
- Serverless and event-driven by default (API Gateway, Lambda, Step Functions, EventBridge, SQS) so a slow model call never blocks a user; ECS/Fargate for long-running work.
- Model layer
- One thin abstraction over provider APIs: versioned prompts, structured outputs, retries, caching, a router that moves a task to another model without a code change.
- Deployment
- Terraform or CDK for every resource, CI/CD through environments, secrets in a managed store, rollback as a single deploy.
- Monitoring
- A trace per request across retrieval, model calls, tools and writes; cost telemetry by feature and tenant; CloudWatch alarms on latency, errors, spend.
- Evaluation
- A golden dataset built with your domain experts before tuning begins, a regression suite in CI, scoring calibrated against human graders.
- Security
- Least-privilege IAM, tenant isolation at the data layer, an audit trail of model and tool calls, prompt-injection defences at untrusted boundaries, redaction wherever personal data would otherwise reach a provider.
Model providers are components, not the product
We do not train foundation models and we do not sell a favourite one. The model is a part selected per task, like a database or a queue.
OpenAI, Anthropic, Google, models on AWS Bedrock and open-weight models are all candidates, chosen on evidence from your workload, task by task. A residency rule can decide the choice as firmly as a benchmark; fine-tuning or a small task model earns a place only where a narrow task justifies it. Because the choice will change, the model layer above lets a provider be swapped and the evaluation suite says whether the replacement is as good. When a better or cheaper model ships, you move; where uptime justifies a second provider, the router carries failover.
| Criterion | What we test on your workload | How it shapes the build |
|---|---|---|
| Quality | Accuracy, schema compliance and refusal rate on your golden dataset, per task. | The best model per task, not one for everything. |
| Latency | Time to first token and end-to-end response under realistic concurrency. | Streaming where a person waits; queues and batching where nobody does. |
| Cost | Tokens per request across the real spread of inputs, projected to volume. | Routing, caching and context trimming so spend tracks usage. |
| Data residency | Where each provider processes and retains prompts, and which shortlisted models are callable from the region you need at build time. | Storage and orchestration deployed in the region you need; for Australian data residency that is the AWS Sydney region. Redaction or an open-weight model where a call must stay onshore. |
| Tooling | Structured outputs, tool calling, long context, prompt caching, batch APIs, and how stable each is. | Only capabilities every shortlisted provider offers, so a swap stays cheap. |
What a typical custom AI application looks like
The boxes change with the brief; the layering rarely does.
- Interfaceshow people and systems reach itWeb app (React/Next.js)Mobile appAdmin panelEmail and document intakePartner API
- Application and APIbusiness logicAPI GatewayLambda handlersECS/Fargate servicesCognito authenticationTenant contextBackground jobs
- Orchestrationprompts, tools, routingStep Functions workflowsVersioned promptsTool definitionsModel routerStructured outputsApproval statesEventBridge and SQS
- Knowledgewhat the model readsIngestion pipelineChunkingEmbeddingsVector index (pgvector)Hybrid searchRerankingPermission filtersPostgres
- Modelsprovider APIsBedrockOpenAIAnthropicGoogleOpen-weight modelsFine-tuned task models
- PlatformAWS, normally your own accountS3RDS/PostgresDynamoDBCloudFrontIAMSecrets storeTerraform/CDKCI/CD
- Observability, evaluation, securitywhy it keeps workingRequest tracingCost telemetryCloudWatch alarmsGolden datasetRegression suiteAudit trailRedactionInjection defences
How a build runs
Our five-stage engagement, as it runs for a build. The AI project cost calculator gives an indicative range before we talk.
- Two free conversationsA 30-minute intro call and a one-hour pre-discovery to confirm the problem is worth a build.
- DiscoveryTwo to four weeks on your data, systems and constraints; use cases scored on impact, effort and risk.
- StrategyTwo to four weeks on architecture, model strategy, build-buy-defer calls and a costed plan the board can sign.
- Build in 12-week blocksVertical slices shipped every couple of weeks to an environment you can use, with evaluation scores alongside.
- HandoverDocumentation, runbooks and dashboards. Keep us on, agree a support period, or take it in-house.
Four patterns we build
Each has a human decision point where the stakes call for one.
Document intake and extraction
Invoices, referrals, applications and contracts arrive by email or upload; the system classifies each, extracts fields into a schema, validates them against your records and queues exceptions for a person.
Classification and routing
Email, tickets, leads and documents sent to the right queue, team or workflow, with the model's reasoning stored so a supervisor can audit and correct it.
Forecasting and prediction
Demand, churn, delays and risk scored from your own history. A classical model often beats a language model here; our machine learning consulting covers when that holds.
Customer-facing copilots with guardrails
Assistants that answer from approved content, take a bounded set of actions, hand off to a person on anything sensitive and log every turn. Guardrails live in code, not the prompt.
Custom AI solutions we have shipped
Client-reported outcomes from three production systems; more in our case studies.
DisabilityAssessments
Automated matching and alerting, end-to-end invoice processing for providers and plan managers, automated reconciliations, secure patient-data storage. 90% of operations automated.
Read the DisabilityAssessments case studyModedu
Custom SaaS platform and admin panel, agentic workflows across operations, AI that produces personalised academic papers. 2× revenue within six months; $400k to $1M ARR (client-reported).
Read the Modedu case studyLuxpip
Web and mobile apps, admin panel, custom B2B API and agentic solutions for internal operations. Manual workload down 75%; scale-up of 40%.
Read the Luxpip case studyThe team that builds it also runs it
One team, brief to production
A small senior team of AI developers, designers and data engineers covers strategy through delivery, remote-first across Australia, onsite when it helps.
Built to be yours
Repositories and infrastructure-as-code are yours from day one of any paid stage, in your own AWS account, or one we set up for you, so your engineers can take over at any point.
Priced per stage, exit at any boundary
Discovery from a fixed fee, Strategy priced upfront, Implementation per 12-week block. No mandatory retainers.
Handover is part of the job
Documentation, runbooks and handover are included; taking the system in-house is a standard option, not a negotiation.
Adjacent work we do
AI engineering
Retrieval, routing, evaluation and observability in production.
Harden a system already in productionAI agent development
Agents that plan, act and stop for approval where it matters.
Custom AI agentsAI integration
Model outputs wired into Xero, MYOB, HubSpot, Salesforce and internal systems.
AI integration servicesAI product development
AI-native SaaS products from validated idea to paying customers.
AI SaaS product developmentAI consulting
Which use cases to fund, in what order, and what to buy rather than build.
AI consulting in AustraliaAI automation
Agentic workflow automation through everyday operations.
Agentic workflow automationQuestions buyers ask before a build
Anything particular to your systems gets answered in the free intro call. Comparing vendors first? Read how to evaluate an AI development company.
Book a free intro callWhat can you build in a 12-week block?
Enough to run in production: one end-to-end use case, such as document intake or a knowledge assistant, with interface, orchestration, evaluation, monitoring and deployment around it. Working software reaches you every couple of weeks from the first sprint.
Do we need our own data before we start?
Access to the documents, records and systems the application will work with, not a labelled dataset or a data science team. Discovery checks what exists and how clean it is; the evaluation set is built with your experts during the block.
Can our data stay in Australia, or run on-premises?
Storage, databases, vector index and orchestration are deployed in the region you need; for Australian data residency that is the AWS Sydney region, in your own AWS account or one we set up for you. Where residency requires it, model calls go to a provider that serves the model you need from an Australian region; otherwise we redact around the call or run an open-weight model in your environment. We build on AWS rather than on-premises hardware.
How do you test quality before and after launch?
With the golden dataset built with your subject-matter experts before tuning starts: real inputs, expected outputs, a rubric. It becomes a regression suite that runs in CI on every prompt, model or retrieval change and blocks a release when scores fall; after launch, traces and reviewer overrides feed the same dataset.
What happens after launch?
Handover includes documentation, runbooks and dashboards, plus a 14-day support window after launch, then paid support if you want it, priced in writing like any other stage. Beyond that you choose: keep us on in further 12-week blocks, or take it in-house, which the code and documentation are written to allow. No mandatory retainers.
Tell us what you want built
Bring the process, the product or the idea. The first two conversations are free, and every stage after them is priced in writing before it begins.