Cherchye Leon

For this project, I built a home server from scratch using an old computer. I installed Ubuntu and configured several services. I set up this website, a test website, my personal free Google Drive, and multiple N8N automation workflows. I configured everything with docker to work together seamlessly and ensured that the server was secure, modular and reliable.
First I build it just to run N8N workflows 24/7, now it is an hobby to expand it with new services and functionalities. It was a great learning experience in server management, networking, and dockerization.
The left picture shows a chatbot I made with N8N. This chatbot helps you with any questions you have about Paperbox’s product documentation.

My master’s thesis "A Modular Translation Layer for Static and Deterministic Payload Translation Between Heterogeneous Imperative-Based Agents Using LLMs" focused on developing a model to automate translation between two agents. This was aimed at agents that communicate in machine languages, such as JSON or XML. The goal was to correctly link the structure and content of messages to their representations in each agent, enabling automatic and deterministic translation.
The model I developed uses Large Language Models (LLMs), which are provided with the agents’ behavior files. By leveraging the LLMs’ coding and linguistic knowledge, the model can extract possible message contents and link them between the agents. While the model is not yet perfect, it served as an interesting starting point to explore the possibilities, implementation, and limitations of LLMs.
Below, you can find my paper and presentation.

This was our bachelor project, "Coturnix", where we had to make an existing environment smart. We chose a quail coop as our environment, inspired by my project partner’s hobby. The goal of the project was to design and implement a fully working system. We monitored everything in the coop, processed the data, sent it to Home Assistant, created automations, and built a connected website dashboard. With this project, we also won the "Pitch your STEM Project 2024" award at our campus for best project. We won this prize by creating an Instagram video of our project, which you can watch here.

This was a school project where we had to train an AI model to play the game Pylos. We started with the Minimax search algorithm and later expanded it into an artificial neural network. The project gave us hands-on experience in understanding how AI learns and what factors influence its performance. We discovered how crucial data quality is, and how much can be gained from processing existing data in smart ways. By adding meaningful features, the model’s performance improved significantly. We also saw how the volume of data and the training time strongly impact the outcome, and we experimented with different hyperparameters to observe their effects. Overall, it was a very valuable project that taught us both the fundamentals of AI training techniques and how to apply them in practice. It was also exciting to see the final results and watch our model play the game. Spoiler: It was better than me in the end.

For this project, my partner and I programmed a neural network from scratch in NVIDIA CUDA. We built the neurons, layers, forward propagation, backpropagation, and the entire training process ourselves, without using existing libraries. It was a real challenge, but once it worked, it was very rewarding to see the results. This project gave me a deeper understanding of the fundamentals and details of how neural networks function and are trained. We trained the model on the MNIST dataset, and it was fascinating to watch how the network learned from the data. The most interesting part was implementing everything in CUDA, learning the language, and applying techniques such as striding. We also experimented with performance optimizations, including coalesced memory access and CUDA streams.

For this project, we had to write a Java program that creates a schedule for moving boxes in a warehouse. This was done by simulating the warehouse with the vehicles, boxes, time, and stacks. It was a very interesting challenge because there were many constraints, and the order of tasks played an important role. Everything had to be programmed completely from scratch, which made it both challenging and a real brainteaser. The first step was to build the entire environment and implement all the constraints. After that, we designed and implemented the planning logic. Although it may sound simple in a few sentences, creating the environment and making sure everything worked correctly was a demanding task and a great learning experience. I am very proud of our result, as the generated schedules scored very well in terms of efficiency, correctness, and processing time.