We help you develop products using machine learning

We are ThinkCDS, a Perth based machine learning consultancy. We help you develop products that require machine learning, and our focus is on better control systems.

Why us?

We've worked with

ThinkCDS logo Why ThinkCDS.com ?

ThinkCDS is a machine learning consultancy based in Perth, Australia. Our mission is to make artificial intelligence straightforward to assess and use in your business.

Sample projects

Leak detection from satellite imagery.

Leaks cost Western Australia about $300 million a year, but we can only inspect 10% of the pipe network a year. In 2016 we were funded by a SPUR grant from Landgate with the goal of trying to detect water leaks using the infrared band in satellite images.

The result was a promising proof of concept, reaching the limits of the data available. This is an open project with full source code available on the github page.

3DSmartReclaim - Better control systems with reinforcement learning.

With 3DImageAutomation.com.au in 2018 we developed a method to control a reclaimer's slew with much better performance than existing PID controllers. Typical PID controllers must take a conservative approach because they can't look ahead or fit to complex target curves. Our smart control system increases reclaim rate by more than 10% with a much lower risk of halting downstream processing by overshooting its target rate.

This increased performance is due to the combination of 3D Image Automation's reclaimer software and the application of reinforcement learning algorithms. These algorithms are a general method for applying machine learning to solve control tasks and have a huge potential for a range of control tasks.

Reinforcement learning cryptocurrency trading

A project with spica.ai in 2017 to train crypocurrencies on Poloniex using reinforcement learning. This project led to a reasonable return rate as well as some interesting findings such as a) we had to retrain models every few months to keep them relevant and b) high volume and low volume environments have very different requirements. An earlier draft of our code is available on github.

Our talks at the Perth Machine Learning Group

You can often find us at the Perth Machine Learning Group where we share our work. Here are the materials for some of our past talks: "Machine learning for large point clouds", "When good models go bad", "Beating sonic and knuckles" and "Data augmentation". The slides are available in the comments section of each of the linked events.


Some of our projects are just for fun :).

Visualising the loss landscape in reinforcement learning

When you solve a machine learning project, your optimiser explores to find the point with the lowest error. What does this look like? We mapped the errors as a landscape.
For context, imagine that you are blind-folded on a golf course and you have to find the lowest point using your feet. This is analogous to solving a machine learning problem. You would take a path very similar to the contour map above - starting at a high point (a red contour) and exploring to find a valley where the error is minimized (blue contours).

Teaching sonic to run

As part of the OpenAI retro competition in 2018 we trained a artificial agent to play sonic the hedgehog using a novel implementation of curiosity. For more information see the slides from our talk at the Perth Machine Learning Group

Teaching a ragdoll to walk with a control system.

This interactive demo shows a 2d ragdoll that was trained to move upright. We created the ragdoll using a simple physics system and a human-like range of motion, then we used reinforcement learning to give it a control sytem.

The ragdoll must learn to keep it's head above it's feet and move to the right. This is the result after 44 hours of offline training.

Click the live demo to simulate a "disruption" that the system must robustly respond to. A full screen page is available here


ThinkCDS is a machine learning consultancy based in Perth, Australia. Our mission is to make machine learning straightforward to assess and use in your business. We have experience in time series analysis, medical imaging, and text, and we specialise in:

Aerial imagery

Image processing is one area where machine learning really excels. You may need a lot of data, but we can help you highlight (segment), circle (detect), or classify your satellite, drone, or aerial data to produce information that is valuable to your customers.

Point clouds

Just like image data, point clouds can be manipulated using machine learning with state of the art results. Sometimes conventional methods such as the point clouds library won't work on your laser, drone, or lidar data. Machine learning has the potential to produce more robust and consistent results with less human input. However this is a new and changing application so the first step is to check to see if your requirements, such as data resolution and response time, make this approach suitable.

Reinforcement learning

Our favourite area! This applies machine learning to control systems. It's difficult to learn to drive a car by watching videos, and the same goes for machine learning agents. That's where the field of reinforcement learning comes in. We let the agent learn to drive in a simulator, then transfer the knowledge to the real world. This is a developing field which currently can be applied to control simple devices, even if the environment is complex and non-linear.

Meet our core Team

We have a wealth of experience in delivering machine learning solutions all the way to deployment. We specialize in applying machine learning to point cloud, satellite imagery, and reinforcement learning applications.

Mike's face

Michael Clark

With a Masters in Petroleum Geoscience, Mike has worked in the oil industry before leaving for the data centers of machine learning. He has experience in taking machine learning models all the way to deployment.

Anthony's face

Anthony Dipofi

With more than a decade of programming experience, and having worked as a researcher in the field of chaotic systems and symbolic dynamics, he now uses his considerable investigative skills in the design and development of reinforcement learning solutions as our Chief Data Scientist.

Ivana's face

Ivana S

A developer with a background in statistics. She prefers react over angular and prefers Bayesian rather than frequentist approaches.

Get in touch

We enjoy hearing about interesting business problems. Get in contact us if you would like to organize a no obligations call to discuss applying machine learning to your business problem.

St Georges Tce
6000 Perth
Western Australia