Ing. Gert Mathyssen

Neural networks (TF) + feature extraction

---

We worked together with Gert for six months. During this time, I’ve been impressed by his ability to pick up new tasks quickly and independently.
Gert is diligent and was able to grasp new projects both from a technical and business perspective.
He helped us by building neural networks to process our customer's bank transactions, he also did feature extraction for our new credit rating model.
On a personal level, Gert had become a respected member of the team, liked for his enthusiasm and direct way of communicating.
I would definitely recommend him to anyone needing help in the machine learning field, and I wish him all the best for the future.

Hendrik Brackmann Head of Data Science at Finiata

Close

ML deduplication core with Tensorflow

---

We worked together with Gert to implement a new core for our catalog deduplication service. Not only did he quickly understood the business requirements, but he was also able to deliver what we agreed respecting both our time and quality constraints.
I particularly appreciated the fact that he works independently, but at the same time can successfully collaborate and communicate with a team of developers.
He is professional and committed to what he does, and I am looking forward to collaborating with him again. I would definitely recommend him to anyone who needs a friendly, communicative and competent data scientist.

Nicolò Pignatelli Director of Engineering @ GFG eCommerce Technology

Close

Neural network for object recognition

---

A Belgium robotics company hired me to solve a problem: Crates filled with toys are coming by on a conveyor belt. Because the crates will go into an automated storage unit, they need to be labelled wih what is inside.
This is where my solution comes in. I have built a deep convolutional neural network (DCN) with Tensorflow and trained it with different pictures of the items.
On the factory flow a simple PC takes a picture of the contents of the crate. By running the DCN with this picture (inference) we can determine what items are in that crate.

Below, you can find a demo video I have made for the customer as a proof of concept. The demo only works with five different items. The final neural network in production can distinguish between 25 different items.

Close

Market Magnify

---

This is one of my bigger projects and I did everything myself.
This includes writing the:

• search algorithm to find the events in the big dataset
• Python back end with PortgresSQL database
• front end in HTML and javascript with Django
• Bokeh visualisation and the Bokeh server behind it

I built this personal project as a potential cash cow project, but unfortunately I had to shut it down because there was not enough traction. The connection with the stock exchange has been cancelled, so there is no new real-time data anymore.
You can visit the website here.

Close

Survey data analysis

---

The Flying Health Incubator hired me to do analytics on survey data they collected earlier. After cleaning and prepping the data I made a pandas dataframe and a correlation matrix. The correlation matrix gave a very nice overview, and it made several trends visible. After choosing the most interesting correlations with the management of the Flying Health Incubator, I made dynamic graphs for the most interesting correlations. (Note: the data in the graphs is random.)

Close

Sport data visualisation

---

This was a small project. A customer had a dataset of a number of sports games, and needed some data visualisation. I’ve used a Jupiter Notebook to do the data wrangling and prepping for the visualisations below. The library used to do the visualisations was Seaborn.

Close

Air sensor calculations

---

On Upwork a private person asked me if I could automate the data processing of an air sensor. I was happy to do this job and it ended up being a very fun one because I could combine my engineering and programming skills. Because the customer needed the option to tweak and change my formulas in production, we decided to work with a Jupiter Notebook. The graphs below were the results after the calculations.

Close