Welcome · Bienvenue · Ne y waoogo · I bisimila · Akwaba · Ẹ káàbọ̀ · Barka da zuwa · Karibu · Boyei malamu · أهلاً وسهلاً
I build systems that read, classify and understand language: Byzantine seals, French online hate speech, 17th-century printed books, museum knowledge. Trained in linguistics, I do NLP with the tools of deep learning and the rigour of linguistic annotation.
The problem. A Byzantine monogram superimposes several Greek letters into a single design. Reading one takes an expert, and there are tens of thousands of seals.
The trap. The corpus has 42,660 rows but only 7,499 distinct texts: without deduplication, the test set simply repeats the training set. After a data audit, a weighted loss and a threshold tuned on validation data, CamemBERT reaches an F1 of 0.83 on the “hateful” class.
An app that makes computational linguistics accessible with no prerequisites: 7 tracks from tokenisation to LLMs, an interactive lab and Python running in the browser. One module is devoted to African languages.
Fusing tabular seismic data and audio signals to detect earthquakes and tsunamis automatically. Multimodal fusion outperforms the best unimodal approaches (Random Forest: 88.9% accuracy).
An OWL 2 ontology that does more than store data: the reasoner infers each object's dynasty, period, royal status and current location. 6 SWRL rules, exception handling, consistency checking.
Search by meaning rather than keywords: Sentence-Transformers embeddings, a FAISS index and a Streamlit interface.
I publish Mots & Pouvoir (“Words & Power”), a French-language publication on language, power, philosophy and African thought. Writing about what words do to people feeds what I do to words with models.
I am looking for a 6-month end-of-studies internship in NLP / LLM engineering, starting March 2027, ideally in R&D. Based in Paris, open to relocation.