Maxence Debes
Data Scientist · ML Engineer
About me
I am Maxence Debes. My profile is deeply rooted in Data Science, with a strong specialization in Machine Learning and AI, built on a solid academic background in Mathematics and Computer Science. I enjoy solving complex technical problems by combining mathematical rigor with advanced Python programming to build robust, data-driven solutions.
Work experience
Skills
Programming languages
Python
Java
CypherML & Deep Learning
PyTorch
PyG
Scikit-learn
PEFTLLMs & MLOps
Unsloth
vLLM
Ollama
W&BTools
Git
Docker
Bash
Linux
LaTeXData analysis
NumPy
SciPy
Pandas
MatplotlibSpoken languages
Education
My projects
A mix of academic research, personal builds, and professional work, all centered around AI, data, and software engineering. Some explore ideas I was simply curious about, others solve problems I ran into along the way. Most link to a repository if you want to dig into the code, and a few are still evolving as I keep iterating on them.
Hi! PARIS Hi!ckathon 2025 (2nd place)
Ranked 2nd out of 80+ teams in the 2025 edition of the Hi!Paris Data Science hackathon. Task: Predicting student PISA scores from complex socio-economic backgrounds. Engineered a gated XGBoost model: one regressor estimates the probability of a non-zero score, a second one, trained on positive scores only, predicts its value, and their product reaches an R² of 0.79 on the test set.
View on GitHub
Molecular Graph Captioning (Kaggle competition)
Development of a retrieval-based architecture using contrastive learning to align molecular graph embeddings with natural language descriptions in a joint latent space. Implementation of a dedicated Graph Neural Network encoder using PyTorch Geometric, trained with a contrastive objective and evaluated with BERTScore and BLEU-4.
View on GitHub
Streaming Safety Classifier for LLMs — Capstone Project with DragonLLM (now OVHai LLM)
Reproduced the toxicity classifier from the Qwen3Guard technical report: built a dataset of 24K toxic and safe conversations synthesized with Qwen3-30B from RealToxicityPrompts, with the position of the first toxic token annotated by Qwen3-30B-Instruct, then trained a lightweight per-token safety head on top of a frozen Qwen backbone for real-time, streaming unsafe-content detection.
View on GitHub