AI / ML Engineer
I build AI systems from the ground up mixing knowledge graphs, neural architectures, computer vision and get them out of the notebook.
Final-year M2 Data Science & AI @ Université Paris-Saclay · AI Engineer Intern @ Dassault Systèmes.
about
Currently: Searching work · exploring Teyvat · Doja Cat on repeat !
experience
Built an end-to-end GraphRAG assistant over a skills ontology (ESCO/SKOS): natural-language to SPARQL, an ontology embedding pipeline in Java with an in-memory vector store, and a conversational panel inside the 3DEXPERIENCE Ontology Explorer.
Built a multi-agent RAG assistant for internal knowledge access doing semantic search over 20K+ documents with BGE multilingual embeddings, FAISS indexing, and a GPT model for answer generation.
Built a document-clustering pipeline (TF-IDF + K-Means, scikit-learn, spaCy) over 500+ documents, surfacing recurring topics and the most-referenced resources.
Developed a Variational Recurrent Autoencoder in TensorFlow to detect anomalies in clients' historical records, reaching 94% accuracy and cutting manual data-cleaning effort.
skills
projects
Knowledge graphs, neural architectures, computer vision, applied risk modelling, representation learning each one built rather than assembled from a tutorial.
Ask a question in plain language and get an answer straight from a graph. An LLM turns the question into SPARQL against a hand-modelled RDF ontology, the query runs locally, and a second pass turns the raw bindings back into a sentence. Simplified implementation inspired by the assistant I built at Dassault Systèmes, rebuilt from scratch at a scale you can actually poke at.
The interactive demo lands here, ask a question in plain language, watch the SPARQL get generated and run against the graph.
The problem : a model that says "tumor, 95%" isn't safe to trust on its own; a doctor needs to know when it's guessing, so doubtful cases get a second look instead of being trusted blindly.
The solution : every prediction carries its own uncertainty (MC Dropout), and the least-certain scans are handed back to a human. A triage aid, not a diagnosis tool. Grad-CAM confirms it looks at the tumor, not the background.
A full Transformer built in PyTorch with no high-level libraries, multi-head self-attention, positional encoding, and autoregressive text generation, implemented to understand every tensor rather than to call one.
Built at Hack4Future with Société Générale and won first place nationally: an AI-assisted credit workflow chaining document analysis, cross-document consistency checking, and ML fraud detection into a single risk decision.
A study on multi-model feature fusion: combining deep feature extraction with representation learning to push classification performance past what any single backbone reached on its own.
A deployed meal-planning application built on recommendation algorithms and REST APIs, designed around how people actually eat.
An intelligent student assistant bringing personalized recommendations, resource aggregation, and productivity tools into one app.
Engineering thesis: a multi-agent architecture for contextual reasoning and information retrieval, graded with high honors.
off the clock
I like people, trading traditions, and finding out what makes a place itself either it's the food, the customs, the small specifics you only ever learn from someone who grew up there. Most of what I know outside engineering came from a conversation rather than a book.
800m finswimming, national title and also Lifeguard certification. Years of early mornings in a pool teach the one thing that transfers everywhere: showing up on the days you don't feel like it.
Genshin Impact mostly part decompression, part puzzle box. A good game hands you a system whose rules you have to reverse-engineer and an incredible storyline on top of it.
Six years of it. It didn't stick in the end, but the habit of reading something structured and turning it into something that sounds right did.
Disney, cartoons, the whole catalogue. I've made peace with the fact that the inner child isn't going anywhere and honestly, it's good company.
contact
Looking for a full-time AI/ML engineering role from September 2026. If your team works on retrieval, knowledge graphs, or getting models into production, I'd like to hear about it.