Career

Paris, France

• Tutobac (with Kajou and Libraries Without Borders): an offline app for Senegalese students preparing the baccalauréat, delivered on Kajou micro-SD cards so it works without a connection — 3,000+ students and a 40% improvement in pass rates.
• Karibu (with Kajou and Libraries Without Borders): a multilingual French-learning assistant for migrants, in Ukrainian, Arabic, Pashto and English, expanding to Swahili and Kirundi — 20,000+ learners.
• kProf (with Kajou and Libraries Without Borders): an offline pedagogical assistant for teachers planning lessons, delivered on Ideas Cube nano-servers and Kajou micro-SD cards — 70 schools and 500 teachers in Senegal.
• TelcoGraph: enrichment of a telecom knowledge graph, as a use case inside LLMs4EU, the Digital Europe Programme project building open multilingual models for Europe.
• SynthEthics: synthetic training data for moral judgement, built across many ethical frameworks so a model can reason inside plural moral frameworks instead of performing an alignment it has not done.
• French Open Science: small reasoning models (350M–3B) fine-tuned to read a parsed publication and extract funding, infrastructure, software, datasets and clinical trials — models, data and code released openly.
• French Science Commons: French-origin open-access scientific publications, 2007–2026 — 1,248,860 documents (1,189,628 articles, 59,232 theses) from HAL, OpenAlex, journals and institutional repositories, built for reuse and for keeping Francophone science discoverable.

• open science
• moral reasoning
• low-resource languages
• retrieval-augmented generation
• on-device / edge NLP
• model evaluation / benchmarking
• medical NLP
• dataset construction
• educational NLP
• African languages

• Built the first Russian dataset for token-level detection of AI-generated text — 10,000 technology articles with mixed human and AI authorship.
• Ran the comparison across 16 state-of-the-art models and 4 statistical algorithms, Binoculars and GLTR among them.
• A detection pipeline that is more accurate than the baselines and substantially faster at inference.

• AI-generated text detection
• dataset construction
• model benchmarking

Milan, Italy · remote

• Three EU-funded projects, run in parallel.
• A scoring methodology for sensitive content, and the identification of it across text, image, audio and video, built into a risk-assessment system for EU police forces.
• Cross-lingual classification of biased versus unbiased news sources, for misinformation detection.
• Text mining and information extraction over multilingual news, for criminal-network and firearms-trafficking patterns — the predictive models improved prevention metrics by 20%+.

• content moderation
• cross-lingual classification
• misinformation detection
• information extraction

Moscow, Russia

• Code4ML: 2.5M annotated Python snippets from 100,000+ Kaggle notebooks, under a new machine-learning taxonomy of 80+ semantic categories for what the code actually does — open-sourced.
• Statistical analysis of large learner corpora across about 200 syntactic complexity features, to find which of them predict essay quality.

• L2 language acquisition
• programming language processing
• dataset construction
• corpus linguistics
• educational NLP

Moscow, Russia

• Rule-based pattern extraction and sentiment lexicons for Italian — the semantic patterns and the relations between them.
• Cut computational complexity by 60%+ while improving accuracy.

• rule-based NLP
• sentiment lexicons
• information extraction
• Italian

Groningen, Netherlands

• Worked on the SemEval Mu-SHROOM shared task.

Moscow, Russia

• Thesis A Comparison of Text Complexity Assessment by Several Tools Using English Learner Texts as the Material, supervised by Dr. Olga Vinogradova.
• Thesis grade 9 out of 10.

• NLP
• Python
• R
• Databases
• Neural Networks
• Linear Algebra
• Calculus
• Statistics
• Syntax
• Semantics
• Morphology
• Pragmatics
• Formal Methods in Linguistics