Machine learning consulting in Australia
ComplxAI provides machine learning consulting in Australia for organisations with a recurring decision and the history to learn from: what to stock, who will churn, which transaction looks wrong. We audit the data, set a baseline, build the model and run it in your AWS account.
- Models, code and pipelines are yours from the first paid stage
- Data and machine learning is a named role on our senior team
- Each stage priced in writing, with an exit at every boundary
Built for decisions you make every week
Machine learning earns its place where one decision repeats at volume and the answer is hidden in history: reorder quantities, fraud checks, which leads to call first. If a rule of thumb does that job today, a model is the next step.
Language, documents and conversation sit with our generative AI consulting. A whole product around a model is custom AI development. Deciding which problems deserve a model starts with AI consulting. This page covers predictive analytics: forecasting, scoring, detection and optimisation, shipped as running software.
A good fit if…
- You hold history of the outcome you want to predict
- A spreadsheet or rule of thumb makes a decision that costs real money
- A model was built once and nobody knows whether it still works
Three ways these projects stall
The algorithm is rarely the missing piece.
The forecast lives in a spreadsheet
Demand and cash-flow numbers are re-cut by hand each month from exports nobody trusts. We replace the export with a pipeline and the ritual with a scheduled run.
The model nobody can defend
A contractor left a model with a flattering accuracy figure and no baseline. We re-test it against the rule it replaced and keep it only if it wins.
The notebook that never shipped
The experiment worked on a laptop and stopped there because nothing was repeatable. We turn it into pipelines, a registry and a deployment that retrains on schedule.
What our machine learning consulting covers
Data science and ML development by the same people, from first look to production run.
Data audit and feasibility
Is the outcome recorded, how far back, how consistently, and does anything leak the answer? Feasibility in writing, including the case for not proceeding.
Baselines before models
A moving average or last year's figure sets the number to beat. Every model reports its lift over it, or it does not ship.
Model development
Gradient-boosted trees, regression and time-series methods for tabular data; fine-tuned open-weight vision and language models for images and documents. The simplest method that clears the bar.
Evaluation tied to the decision
Accuracy is rarely the metric that matters. We score on the cost of a missed fraud case or a stock-out, on data from after the training period.
MLOps: pipelines to retraining
Feature pipelines and a feature store, a model registry with versions and approvals, drift alarms in CloudWatch with a rollback path, and retraining through the same evaluation gate.
Governance and explainability
Per-prediction explanations for the people acting on the output, lineage of which data trained which version, and human approval where a prediction changes a customer's outcome. Training data stays in your own AWS account, or one we set up for you, deployed in the region you need; for Australian data residency that is the AWS Sydney region.
An ML platform on AWS, layer by layer
The model is one box. The other layers keep it accurate over time.
- Data sourceswhat the model learns fromS3 data lakeRDS/PostgresDynamoDBXero, HubSpot and Salesforce exportsImage and sensor feeds
- Feature layerrepeatable inputsFeature pipelines on Step Functions and LambdaFeature storeQuality checksPoint-in-time joins
- Trainingexperiments to candidatesSageMaker training jobsECS/Fargate containersExperiment trackingTime-ordered validation
- Registrywhat is allowed to shipModel registryBaseline comparisonDecision-cost metricsApproval gateRollback
- Servingbatch and real-timeBatch scoring to Postgres and S3SageMaker endpointsLambda inferenceAPI GatewayEventBridge schedulesWrite-back to source systems
- Monitoringdrift and retrainingInput and prediction driftOutcome feedbackCloudWatch alarmsRetraining triggersTerraform/CDKCI/CD
Batch scoring on a schedule is the default, because most business decisions are daily or weekly. Real-time endpoints on SageMaker or Lambda are added where the decision happens inside a request, sized alongside our AWS consulting so idle capacity stays off the bill.
Predictive models we build
Typical shapes of work. Each is scoped to your data and your decision.
Demand and cash-flow forecasting
Weekly demand by product and location, or cash position by week, with seasonality handled and prediction intervals attached.
Lead and churn scoring
A score per lead or account from CRM, usage and billing history, written nightly into HubSpot or Salesforce.
Anomaly detection
Transactions, claims or sensor readings that break the pattern, flagged with a reason and tuned to reviewer capacity.
Classification and triage
Tickets, applications or inspections routed to the right queue, with uncertain ones sent to a person.
Recommendation
Next product, content or action from behaviour and catalogue data, batched for email or live on the page.
Document and image understanding
Fields extracted from invoices, forms and plans; defects or counts found in photos; fine-tuned open-weight models where a general API falls short.
Perception and decisioning
Models for hardware that must act, not report: perception and decisioning for an agricultural drone startup under NDA that later expanded into the US.
Pricing and optimisation
A model estimates the response, an optimiser picks the price, allocation or schedule, and a person signs off above a threshold.
From data audit to a model that keeps learning
The work runs inside the five-stage engagement in how we work; Implementation blocks deliver these steps.
- Data auditCoverage, quality, leakage and labelling of the outcome. Feasibility in writing.
- BaselineThe rule of thumb or current process, measured properly.
- ModelFeatures, candidate methods and a training pipeline in your repository.
- EvaluateTime-ordered held-out data, scored on the cost of the decision.
- DeployBatch, real-time or edge serving in your AWS account.
- Monitor and retrainDrift alarms, outcome feedback, retraining through the same gate. Handover.
Predictive machine learning versus generative AI
Often bought under one label, they solve different problems. Many systems need both: a small predictive model routing work inside an LLM pipeline is AI engineering. Our LLM consulting covers the right-hand column.
| Aspect | Predictive machine learning | Generative AI (LLMs) |
|---|---|---|
| Question it answers | How much, how likely, which category, which is unusual | What does this say, write this, take the next step |
| Data you need | Enough history of the outcome to learn from | Often little; retrieval over your documents instead |
| How it is judged | Lift over a baseline on time-ordered held-out data | Evaluation sets and human-calibrated judges |
| Where we start | Data audit and baseline | Use-case map and evaluation set |
Consultants who also run the system
A named seat for data and ML
Felix holds the data and machine learning role on the team, next to engineering, infrastructure and design, so a model is never thrown over a wall to be productionised.
Your models, your account
The repository and the AWS account it deploys to are yours from day one of any paid stage. Feature code, training pipelines, model versions and the Terraform behind them are handed over as standard. No retainer attached.
Models and platforms we have shipped
Perception and decisioning models for an autonomous agricultural drone startup, under NDA, that went on to expand into the US. The same team built the platforms behind DisabilityAssessments and Luxpip. See the case studies.
Around the model
AI engineering
Retrieval, routing and evaluation around the model.
Production AI engineeringAI development
Applications built around models, tools to platforms.
AI development servicesAWS consulting
The account and serverless platform the models run on.
AWS architectureGenerative AI
LLM systems for language, documents and conversation.
Generative AI consultingAI consulting
Which problems deserve a model, scored by impact, effort and risk.
AI consulting in AustraliaAI integration
Scores and forecasts written into Xero, HubSpot and Salesforce.
AI integration servicesQuestions we get before a model is built
Bring a data dictionary, or just describe the decision. The first two conversations are free.
Book a free intro callDo we have enough data?
Often more than you think, sometimes the wrong kind. What matters is whether the outcome was recorded for a full business cycle without the answer leaking into the inputs. The data audit settles this in writing before any model is built.
How long does a machine learning project take?
Discovery runs two to four weeks and ends with a use-case map scored by impact, effort and risk. Implementation then runs in 12-week blocks with working software every couple of weeks, each block priced in writing before it starts.
How will we know it works?
By comparing it with what you do now. Every model reports its lift over a baseline on data from a period it never saw, in the units of the decision, re-run at every retrain.
Should we buy a model or build one?
Buy when a vendor's model is trained on data like yours and the decision is generic. Build when your history is the advantage or the output must sit inside your systems. Strategy makes that call per use case.
Should we just use an LLM instead?
For text, documents or conversation, possibly. For how much, how likely or which is unusual, a predictive model on your own history is usually cheaper, faster and easier to explain. Many systems use both.
What does it cost to keep running?
Batch scoring on serverless infrastructure keeps compute cheap; real-time endpoints cost more because something is always on. The larger item is people time for monitoring and retraining, which we automate and hand over. The AI project cost calculator gives a rough build range.
Bring the decision and whatever data you have
We will say whether a model is warranted, what it takes to beat the rule you use today, and what each stage costs.