About
Introduction
Hi, I'm Avrile, a PhD Track Master student in Data & AI at Institut Polytechnique de Paris (École Polytechnique and Télécom Paris).
I work on mechanistic interpretability, AI fairness, and legal NLP. I'm drawn to questions that cut across these: how biases arise in language models and how to measure them, how far interpretability can be trusted, and how NLP interacts with the law and the regulatory frameworks that govern AI.
- Trajectory: law (LL.M. Duke · financial law Paris-Dauphine · banking compliance at PwC) → teaching French as a foreign language → AI/NLP at IP Paris
- I read and write in French and English, and speak Spanish
- Outside the lab: PADI Divemaster, blackwater and pelagic diving
Research lines
Click any entry for detailsInterpretability Auditing for Bias Measurement
When can readouts of a model's internal representations be trusted for bias measurement? Joint research on evaluation protocols for the reliability of interpretability methods in social-bias auditing, across multiple open-weight LLMs.
Read more →Argument Mining on French Supreme Court Rulings
A large-scale extraction pipeline exploiting the codified rhetorical structure of French cassation rulings to map each ground of appeal to the court's response: 121K+ argument pairs across 86K decisions and the five civil chambers. Lead author of the resulting dataset and tool paper, with Tamara Dhorasoo, Nils Holzenberger, Thomas Le Goff, Tiphaine Viard, and Maria Boritchev.
Read more →Cultural Binding Heads
Why do LLMs default to equal treatment across cultural groups even when context warrants differentiation? I identify the 2–3 mid-layer attention heads that causally mediate cultural binding across eight language models, and find that models know 3–5× more than they act upon: the bottleneck is routing, not knowledge. Lead author of the resulting paper, with Luca Benedetto.
Read more →Where Experts Disagree, Models Fail
A benchmark of 1,015 passage–article pairs in French civil law, annotated by three legal experts (κ = 0.33). Annotator disagreement turns out to be the strongest predictor of model failure on implicit statutory citation. Lead author of the resulting paper, with Tamara Dhorasoo, Soline Pellez, and Nils Holzenberger.
Read more →Publications
Selected writingCivil
Cassation
A Grammar of French Civil Cassation: A Ground-Level Dataset and Tool
Awards
RecognitionIP Paris PhD Track Excellence Scholarship
Competitive merit-based scholarship from Institut Polytechnique de Paris: a cost-of-living grant plus subsidized tuition across the two years of the Master's.
Kaggle: Influencer or Observer · 1st of 87 teams
Winning solution for the Deep Learning course (INF554) Kaggle competition. NLP classifier predicting users' social roles from tweet data alone.
Projects
Selected technical workInfluencer or Observer: Predicting Social Roles on Twitter
1st of 87 · 88.0% accuracyWinning entry for the INF554 Deep Learning Kaggle challenge at École Polytechnique, with Falguny Barua Ema and Saurabh Mishra. A stacking ensemble that predicts whether a Twitter user is an influencer or an observer: LightGBM over 1,373 user-level behavioral features, fine-tuned CamemBERTa-v2 reading multi-view "user cards" of aggregated tweets, and a HistGradientBoosting meta-learner combining both signals. Behavioral and linguistic features turned out to be complementary, which is what pushed the ensemble past either model alone.
How many dogs to herd sheep?
NetLogo agent-based study, with Prof. Ada Diaconescu (LTCI, Télécom Paris), of how herding effort scales with flock size. Proposes a local sheep-density gradient as a tractable proxy for the flock's centre of mass.
Repo ↗French electricity consumption forecasting
Interactive R Shiny app for rolling forecasts of national electricity demand, with built-in model performance testing.
Live demo ↗Self-Organizing Map classifier
A Kohonen self-organizing map implemented from scratch in C, used as an unsupervised classifier.
Repo ↗Education
TrajectoryPhD Track Master in Data & AI ranked 1st of 30 · 2025–2026
Advisor: Prof. Nils Holzenberger. Selected coursework: Deep Learning · Collective Intelligence · AI Ethics · Logic & Knowledge Representation · DBMS · Probabilities.
Bachelor's in Computer Science summa cum laude
Selected coursework: Algorithms & Data Structures I–II · Interpretation & Compilation · Operating Systems · Networks · Data Mining · Language Engineering.
Mathematics & Statistics 30 ECTS
Linear Models · Time-Series Modeling & Forecasting · Statistical Techniques · Scientific Foundations of Mathematics.
Master's in Teaching French as a Foreign, Second, or Specific-Purpose Language cum laude
Master's in Financial Law (M2) cum laude
Master of Laws (LL.M.) GPA 3.226/4
Certificate in Business Law.
Bachelor's in Law
University Diploma · Comparative & International Legal Studies cum laude
Teaching
PedagogyAcademic Tutor · Computer Science Bachelor's Program
Academic and methodological support for undergraduate Computer Science students; liaison between teaching staff and students.
French Teacher
Six years of one-on-one and group instruction across all CEFR levels (A1 to C2); personalised curricula and weekly teaching materials tailored to each learner; an automated note-organisation tool to streamline lesson follow-up.
Contact
Get in touchThe easiest way to reach me is by email. Don't hesitate to get in touch.