Brisbane AI at night

The Current Status of the Brisbane AI scene

Brisbane has always been the home of good weather, great beaches close to the city and an enviable outdoor lifestyle, however, now it has something new – a blossoming Artificial Intelligence (AI) scene.

From the outside sunshine, now to the inside monitor glow, read on to discover some exclusive insights on the Queensland capital’s thriving ecosystem. 

Brisbane AI

With recent investments by the Queensland government in Brisbane-based AI initiatives, like the AI Hub ($3M as part of a larger $650M commitment), there has been significant growth in the number of companies: using AI, building products and services based upon the new technology, or simply consulting in the field.

To understand how big this growth has been, Biarri has gone out and put its finger on the pulse of the burgeoning industry to understand how far we’ve come over the last decade and the direction of Brisbane’s future trajectory – based on the graphs below, the numbers are phenomenal. For example, we’ve seen over 2500% increase in the number of people employed in companies working with artificial intelligence or similar technologies (from around 20 to 500+).

To carry out our analysis we approached our Brisbane networks, including key individuals active in the AI scene, to reach out to their AI contacts and let us know the current state of the industry. We then cross-referenced this with LinkedIn, job boards such as SEEK and Indeed, as well as various websites to ensure the data was as accurate as possible (at the time of writing). Based on this we were able to put together a comprehensive list on the state of AI-ffairs in Brisbane.

However, before we present the data showing the surprising results, we asked ourselves: what has driven this growth in activity around machine learning and artificial intelligence?

We hypothesise that there are four main factors that we discuss in the following:

  • Enhanced ease-of-use of tools to build AI solutions
  • Improvements in the power of the available AI technology
  • Increased number of people with AI skills
  • Increased awareness and interest in AI


In the last 10 years we have seen the barrier to entry for using AI tools reduce considerably. This has been driven largely by a strong open-source community and cloud-based tools that have led to:

  • Orders of magnitude cost reduction for at-scale AI deployments
  • The commoditisation of cutting edge AI algorithms such as neural nets, boosted decision trees, etc.
  • Programmable infrastructure for the deployment of web based applications
  • Simple APIs providing access to leading open-source machine learning and artificial intelligence libraries such as TensorFlow, PyTorch, MXNet, etc.

Improvement in power

In addition to the lower barrier to entry, we have witnessed an unprecedented improvement in the quality of algorithms powering AI inferences. The new algorithms are able to converge to solutions faster with less training in much smarter ways (for example pretrained models).

In addition, the cloud architectures available now provide the ability to store and analyse massive data sets, further improving the quality of these algorithms. The storage of a Terabyte of data in the cloud costs on the order of $20 per month with a number of providers.

Growing talent pools

Almost unlike any other area, online courses for machine learning and artificial intelligence have exploded. Massive Online Open Courses (MOOCs) offered by edX, Coursera, and Udacity provide courses delivered by leading AI experts – many of which are completely free. There are also countless large, well-known, and high-quality blogs providing up-to-date information on AI.

Beyond that, universities around the globe have caught onto the trend and are offering Masters in Data Science for students to cater to the demand. Many of the courses are quite new, for example the one at the University of Queensland in Brisbane is less than 2 years old at the time of writing.

Looking further abroad, Georgia Tech have teamed up with Udacity and AT&T to offer the first accredited Online Master of Science in Computer Science (OMS CS) that students can earn exclusively through the MOOC delivery format. This is changing the landscape of education globally.

Increased awareness and interest

Another big contributor is the general interest in AI. Looking at Google Trends, the following graph shows the number of searches for AI-related terms over the last 10 years. The graph shows an up to 10x increase in interest artificial intelligence, machine learning, and data science in the last five years, so we are truly entering a period of increasing awareness and interest.

What effect has this had on the Brisbane AI market?

The effect of the above has been noticed in all sectors. The following graphs show what we discovered.

If we just examine companies offering AI services, this first graph shows the simple growth in the number of companies:

We can see the hockey curve style growth starting to kick off in 2014. However, where is this growth occuring?

To figure this out we need to better understand the types of companies operating in this space. To help understand the types of companies that have grown out of the trend, we split them into the following categories:

  1. SME that provides consulting
  2. Non robotics AI startup
  3. Incubator
  4. Robotics AI startup
  5. Larger company that provides mostly AI consulting

In the following graph we show the growth of such companies with the above split:

We can see that the growth has largely been due to robotics and non robotics AI startups.

Naturally this proliferation of AI companies has led to a large increase in the number of people employed in companies driving the AI agenda as the following graph indicates:

The number of people employed in these areas is being felt beyond employment with Meetup groups like Queensland AI (2,900 members) and Brisbane Data Science (1,800 members) thriving in recent years.

Another interesting development has been the congregation of such companies, leading to the government’s choice of the location of the AI Hub to be a prudent one. There is a strong collection of such companies around the Fortitude Valley as the following heatmap of Brisbane demonstrates:

Number of Brisbane AI companies by location in Brisbane

The colour coding is such that there are now 11 companies in the valley, 7 in the city and elsewhere either 1 or 2.

So if you want to join the blossoming AI scene then come and join us down in the Valley.

What’s next?

As can be seen from the above, the current AI scene is blossoming with more and more companies entering the artificial intelligence fray. If the numbers continue as above (although bottlenecks like talent shortages will most likely arise soon) then by extrapolating the data with an exponential curve, we could see a doubling of the number of people working in AI by the end of 2023 to over 1000 people.

So although the industry started slowly we expect it now to now have doubled within four years.

This bears well for the graduates coming out of universities but also to service providers to the AI industry as it would be expected that similar trends exist worldwide. With a wealth of resources already existing online to help people enter the industry, we only expect speed of adoption of AI to increase.

Get your hands on this data

To access the data, click on the link here to view the spreadsheet with the information that we have collected.

But in the meantine, what are your predictions? Did we miss anything?

Get in touch

  • This field is for validation purposes and should be left unchanged.

SaaS Optimisation vs Traditional Approaches

SaaS Optimisation solutions have become a necessity in modern business. These solutions don’t just result in lower costs, better products, and ease of use, but better customer satisfaction, employee satisfaction, and streamlined business processes.

Gartner predicts the SaaS market to be $85.1 billion globally by 2019, up from $72.2 billion in 2018. While much of this growth will be the continued adoption of infrastructure such as CRM and ERP, point solutions will continue to emerge to fill gaps.

What are point solutions and why should they matter?

Unlike ERP systems that provide end-to-end solutions for problems that you might not have, point solutions are targeted to specific pain points within your business. This might be: The best way to route your vehicles, the best way to manage your employees, or, the best way to build your facilities (location and size).

Point solutions also give you the capabilities of integrating with larger ERP systems.

A unique problem requires a unique solution.

The Key Benefits of SaaS

The benefits to a SaaS solution over a traditional approach can be evident by looking at these 5 key areas:

  • Reduced time to benefit. SaaS development reduces the feedback cycle to the point where anyone, anywhere in the world can test software on a web browser.
  • Lower costs. You don’t require expensive IT infrastructure and big IT departments, these worries are outsourced to the company offering the solutions.
  • Scalability and integration. SaaS solutions can be scaled up and down based on your requirements. This means you can drive more powoer in peak times or seasonality.
  • Faster releases (upgrades) as it reduces the feedback loop and allows for much easier troubleshooting.
  • Easy to use and perform proof of concepts

These benefits have allowed Biarri to develop world class tools for operational and strategic business requirements.

SaaS vs Traditional Software at Biarri

Beyond the above generic benefits, there are a number of specific benefits that we see at Biarri. We believe that the SaaS approach is far superior to ensuring the client gets a solution into a production environment sooner rather than later, it also makes the job of support much more efficient as well. Biarri’s specific benefits it sees are outlined in the below table.

SaaSTraditional Software
DevelopmentEasy to get constant feedback using the Mitigation by Iteration approach. More iterations and feedback from clients can significantly reduce development timeframes.The client either needs to wait for the final version to be installed in their testing environment. If iterations are used they are usually far less frequent (and more difficult to manage) than a SaaS approach.
User Acceptance Testing (UAT)All the client needs to do is use a testing environment URL. No UAT specific hardware is required by the client.Often clients need to setup different computers and/or databases for the UAT environment.
DeliveryClient usually only needs to type the production URL into a web browser. Sometimes their firewall needs to be configured to allow access to our servers.Software has to be installed on each machine which requires more resources from IT departments.
Hardware Clients only need access to an internet connected device. It does not need to be powerful. It could be a desktop or mobile device.Each user needs to have computing power, and memory, required to run potentially resource hungry optimisations.
Software CostOur optimisations sometimes use expensive third party tools (like Gurobi) which are on our servers. This cost gets distributed to all our clients in the form of licensing fees.Each user would need to have the expensive 3rd party software installed on their computer which has the potential to massively increase the final price if there are many users.
SupportWe can easily replicate issues because your data and logs are on our server. This often results in issues being replicated, addressed, and a fix delivered all in the same day.Often it can be difficult to even determine if an issue is caused by the software, or the complex IT environment that it may be running in. Complex IT environments may have internal firewall issues, database issues, Citrix issues, authentication issues, timeout issues, etc. If the problem is with the software it is sometimes required to export the client database and transfer it to the vendor (via ftp, or if it’s too large via snail mail). Sometimes replicating the issue can be difficult because the client’s and vendor’s environment are different (e.g. different database versions, different patch versions or operating system versions). This process can often take weeks even for relatively simple issues.

Get in touch

  • This field is for validation purposes and should be left unchanged.

Making it Accessible: Applying the Optimisation Toolkit

At Biarri Optimisation one of our core priorities is to uncover interesting new mathematical problems in industry, build optimisation engines to solve them, and use those products to incubate new Biarri entities.  It is a constant source of amazement at how many disparate contexts that we can apply the tools of our trade: in particular, the tool of Mixed Integer Programming, which is one of the pillars of our discipline of Operations Research. Tools like this live in our optimisation toolkit.

Operations Research (OR) has its origins in World War II, where mathematics and analysis was used to determine what convoy size was most effective for avoiding German U-boat detection, the best paint colour to maximise camouflage for aircraft, and to identify the best trigger depth for aerial-delivered depth charges to maximise U-boat kills.  You can read more about these fascinating applications and more in the excellent book Blackett’s War.

The problems we solve

We tend to think of the problems that we try to solve in the following way.  First, we identify the decision variables that are involved: what elements of an operation do planners and users have the power to change?  Second, we establish the objective: what are planners looking to minimise or maximise?  This might be cost, or some combination of minimising cost with maximising benefits or safety (for example).  Lastly we pinpoint the constraints: what are the real-world limits on resources (people, equipment, budget) that have to be taken into account?

An example is a daily workforce planning problem faced by warehouse and distribution managers in many different industries.  Here there is a “profile” of required work across the day, of various types (e.g. forklift work, truck loading/unloading, racking replenishment, etc), and a set of personnel who have different skills and availabilities to do this work (permanent, casuals, with various possible start times).  How to cover all the work while keeping the total shift cost as low as possible?

The objective in this example is cost; the decision variables are the allowed shifts that can be operated; and the constraint is that all the work must be done by the appropriately skilled personnel.  The extra complexities we might encounter here come about where people can do multiple task types (e.g. they are qualified for both forklifts and manual truck loading), and there may be a changeover time incurred between tasks; they will also require meal and rest breaks, or incur overtime; there is fatigue to take into account (productivity typically starts dropping later in a shift); and equipment and space/congestion constraints.  You can even apply a productivity multiplier if different people are more or less productive at different tasks. Add to this the fact that some of the work can be done well ahead of time (pre-picking for truck loads, for example), which allows the work profile to be “smoothed” over, and you have quite a complex planning task indeed!

When we formulate this as a Mixed Integer Problem (MIP), we aim to solve it for the overall minimum cost.  This overall minimum cost is known as the global optimum, and it may be that, if you look at any one part of the solution in isolation, it might seem “sub-optimal”.  In our workforce planning example, for instance, a person might be allocated a very short shift. But this will make sense in the overall sense of the problem: it might be that there is a small piece of work left over that cannot “fit” into all the other shifts.

Feasible solutions

Part of the art of applying this type of mathematical approach lies in making sure that there is always an implementable solution.  For example, what if there are not enough people to cover the work? We do not want the solver to tell the user that the problem is simply “infeasible” – that there is no plan.  Instead, we build in extra decision variables which model the unallocated work over time, and we give these variables a high artificial cost in our objective (so the solver still tries to cover as much work as possible); now we can also report to the planner how much work is uncovered.

This technique is an example of a more general method which distinguishes between hard and soft constraints.  Hard constraints must always be met; there is no wiggle room.  By contrast, soft constraints (as in the “uncovered work” example) allow you to break a constraint, but incur some penalty for doing so.  In the workforce example, overtime can also be thought of as a soft constraint (once you exceed the regular shift limit, you incur extra cost).

Soft constraints can also be used as a way to “explore” the solutions that are near the optimal solution – there are some trade-offs you might be willing to make in order to save an extra chunk of cost.  You sometimes have to be quite careful when there are several types of soft constraint or extra penalties: for example, allowing very large penalties can obscure the smaller components of the objective. When you have multiple competing objectives, it can also sometimes be hard to explain why the result looks like it does.

Trusting the results

Of course, we must also exercise caution when trusting results of a mathematical process.  Our solvers will make arbitrary decisions when we do not provide guidance – a common instance is where there is a “shallow” cost function, which is where there are several similar solutions with the same or nearly the same cost.  These might be identical from an optimisation point of view: in our workforce context, an example is where several people with the same skills and availability might be able to do the same task: if it makes no difference to the total cost, who should we choose?  Often there is a real-world analog to this problem of “breaking ties”: here, for example, we might choose the employee based on seniority; or to ensure a longer term attractive roster; or a host of other factors.

Biarri has always strived to make optimisation “accessible” – part of which means making our optimisation results implementable in the real world – and in these examples I hope you have seen some of the nuances and complexities that underlie this promise as well as the tools that comprise our optimisation toolkit.

By Andrew Grenfell

Get in touch

  • This field is for validation purposes and should be left unchanged.

BAM2018 Wrap Up