I’ve interviewed about forty graduates from Australian university AI and machine learning programs over the past eighteen months. They can build a neural network from scratch, explain backpropagation clearly, and have strong opinions about attention mechanisms. What most of them can’t do is deploy a model to production, set up a CI/CD pipeline for model updates, or explain how to monitor a deployed model for drift.

This isn’t a criticism of the graduates. They learned what they were taught. It’s a criticism of the curriculum, which in most Australian universities remains heavily skewed toward theory and research at the expense of practical engineering skills.

The Gap

The typical computer science or data science degree at an Australian university covers:

• Mathematics (linear algebra, calculus, probability, statistics)
• Machine learning fundamentals (supervised, unsupervised, reinforcement learning)
• Deep learning architectures (CNNs, RNNs, Transformers)
• Natural language processing and computer vision
• Research methodology and paper reading

What’s typically missing or underweight:

• MLOps and model deployment — how to get a model from a Jupyter notebook to a production API
• Software engineering practices for ML — version control for data and models, testing ML systems, reproducibility
• Infrastructure and compute management — cloud services, GPU allocation, containerisation
• Data engineering — data pipelines, data quality, feature stores
• Monitoring and maintenance — model drift detection, retraining strategies, A/B testing in production
• Ethics and governance in practice — not just philosophical frameworks, but practical tools for bias detection, fairness measurement, and compliance

Why This Matters

The Australian Computer Society estimates that Australia needs roughly 650,000 additional tech workers by 2030. A significant portion of that demand is for AI and ML practitioners. But employers consistently report that fresh graduates need 6-12 months of on-the-job training before they’re productive in production ML roles.

That’s not normal for a professional degree. An accounting graduate can prepare a BAS in their first week. A nursing graduate can work a ward rotation immediately. But an ML graduate often can’t deploy a model without significant mentoring.

The problem isn’t that graduates are poorly educated — they’re very well educated in the theoretical foundations. The problem is that the curriculum was designed when ML was primarily a research discipline, and it hasn’t fully adapted to ML’s transition into an engineering discipline.

What Good Looks Like

A few programs are getting it right, and they share common characteristics:

Carnegie Mellon’s MLOps course requires students to deploy a complete ML system with automated training, testing, monitoring, and retraining. Students experience the full lifecycle, not just the modelling phase.

Stanford’s CS 329S (Machine Learning Systems Design) covers system design decisions — where to put the model, how to handle scaling, how to manage the data pipeline — rather than just how to build the model itself.

In Australia, UNSW’s COMP9444 has started incorporating deployment exercises, and University of Melbourne’s COMP90049 includes practical data engineering components. But these are exceptions, and the coverage is typically one unit within a broader program rather than a foundational thread.

What Should Change

1. Teach the Full Lifecycle

Every ML program should include at least one unit where students take a model from data collection through training, deployment, monitoring, and retraining. This should be a capstone experience, not an elective.

The tools don’t need to be complex. A student who has deployed a simple model to a cloud endpoint using Docker, set up basic monitoring, and experienced a model retraining cycle has learned more practical skills than someone who has optimised a transformer architecture on a benchmark dataset.

2. Require Software Engineering Fundamentals

Git, testing, CI/CD, code review, and documentation are not optional skills for someone building production ML systems. Too many ML graduates treat code as disposable notebook cells rather than maintainable software. Version control for code, data, and models (using tools like DVC or MLflow) should be taught alongside the algorithms.

3. Include Data Engineering

A model is only as good as its data pipeline. Teaching model architecture without teaching data quality, data validation, and feature engineering is like teaching engine design without teaching fuel systems. Students should understand how data flows from source systems into training pipelines and how feature stores work in production.

4. Teach Cost Awareness

Cloud GPU compute is not free, and universities often shield students from this reality by providing pre-provisioned compute resources. Understanding the cost implications of model size, inference optimization, and infrastructure choices is a critical professional skill.

5. Bring Industry Practitioners Into the Classroom

Academic instructors are essential for theory. But for practical skills, experienced practitioners bring perspective that pure academics can’t. Guest lectures, industry mentors, and adjunct instructors from companies actively deploying ML systems would bridge the gap between classroom and workplace.

The Employer’s Role

This isn’t entirely on universities. Employers need to invest in graduate development programs that close the practical skills gap. Expecting a fresh graduate to independently build and deploy a production ML system on day one is unrealistic regardless of how good the curriculum is.

Companies that run structured onboarding — pairing graduates with senior ML engineers, providing sandbox environments for practice, and building internal training programs — get productive team members faster and retain them longer.

Some organisations are also creating their own AI education programs. Internal bootcamps, tool-specific training, and MLOps certification pathways fill the gaps that university programs leave. It’s a workaround rather than a solution, but it’s better than complaining about graduate quality without doing anything about it.

Looking Forward

The good news is that awareness of this gap is growing. Australian universities are beginning to revise their AI curricula, partly in response to employer feedback and partly because the field itself has shifted from research-first to deployment-first. The Australian Government’s National AI Centre has also been pushing for more practically oriented AI training.

The transition will take time. Curriculum changes in universities move slowly — a new unit proposed today might not be taught for 2-3 years. In the meantime, graduates who supplement their degree with practical experience — personal projects, open-source contributions, internships — will have a significant advantage in the job market.

The AI field doesn’t need more people who can write a perfect neural network on a whiteboard. It needs more people who can get that neural network running reliably in production, at scale, within budget. The sooner our education system reflects that, the better off everyone will be.