I'm an AI Engineer at Faktion (Antwerp), building LLM-based and agentic AI solutions for enterprise clients, from prototype to production on Azure.
I graduated Summa Cum Laude from UCLouvain (École Polytechnique de Louvain) with an M.Sc. in Computer Science & Engineering, specializing in Artificial Intelligence. My master's thesis on deep reinforcement learning was presented at IEEE CoG 2026 and won 2nd place at the IEEE/ICTEAM Best Master Thesis Award 2026.
I'm passionate about Generative AI, Agentic AI, Machine Learning and Computer Science as a whole. What I enjoy most is building the entire pipeline end to end, from the agentic core to full-stack delivery, and seeing the result come to life in production.
Happy to connect about applied AI, agents in production and AI research: LinkedIn or mathis.delsart@gmail.com.
|
ML & DL |
Generative AI |
Web & Backend |
Databases |
|
DevOps & Cloud |
Systems & Embedded |
Dev Tools |
LLM Platforms |
Deep Reinforcement Learning for Competitive Agents in MicroRTS: Architecture, Training, and Tournament Evaluation
UCLouvain · June 2026 · Presented at IEEE CoG 2026 (Madrid) · 🥈 2nd place, IEEE/ICTEAM Best Master Thesis Award 2026
A DRL agent for real-time strategy games, fusing a U-Net spatial encoder with an entity-level Transformer (UECD) and trained end-to-end with a modular PPO pipeline on HPC GPU clusters (SLURM), with every design decision individually ablated.
- 🥇 96.67% pool win rate, topping a 19-agent tournament (1st on 4 of 5 metrics)
- ⚔️ 9-1 head-to-head record against RAISocketAI (prior MicroRTS competition winner)
- ⚡ Trained on just 9.4 GPU-days, roughly 7× less compute
UECD-Best (ours, top-left) vs RAISocketAI (bottom-right)
Combining Spatial and Entity-Based Reasoning for Competitive MicroRTS via U-Net and Transformers
Mathis Delsart, Achille Morenville, Éric Piette · IEEE Conference on Games (CoG) 2026, Madrid
BibTeX
@inproceedings{delsart2026uecd,
title = {Combining Spatial and Entity-Based Reasoning for Competitive MicroRTS via U-Net and Transformers},
author = {Delsart, Mathis and Morenville, Achille and Piette, {\'E}ric},
booktitle = {Proceedings of the IEEE Conference on Games (CoG)},
year = {2026},
address = {Madrid, Spain}
}Generative AI & LLMs
| Project | Repo | Description | Stack |
|---|---|---|---|
| Sourcio: RAG Study Assistant | Deployed RAG tutor answering only from your own courses, with citation-by-construction (the model never sees a page number) + a three-layer refusal guard. Live app: 100% citations, 96% refusal accuracy | ||
| NLP-Models | Progressive NLP pipeline: Question Classification (Naive Bayes), Vector Semantics (TF-IDF/PPMI/Word2Vec), BERT-based QA (SQuAD) | ||
| Text-Prediction-TwitOZ | Intelligent text prediction system using N-grams algorithm with real-time word suggestions |





