PhD Candidate in Applied Mathematics, McGill University
Open to full-time research roles, starting Fall 2027I am a PhD candidate in Applied Mathematics at McGill University, advised by David Stephens. My research focuses on Bayesian Deep Learning and Uncertainty Quantification. I expect to complete my PhD in Spring 2027.
I develop scalable variational methods to make neural networks robust, interpretable, and computationally efficient, bridging theoretical rigor and practical deployment to build safe and trustworthy AI systems.
In 2024 I was awarded the Women in AI Excellence Scholarship from Mila (Quebec AI Institute). My academic foundation was built in France: after the intensive Classes PrĂ©paratoires aux Grandes Ăcoles, I earned an Engineering Degree from ENSIIE and an MSc in Quantitative Finance from Paris-Saclay University, where I graduated with First Class Honours and received the Sophie Germain Excellence Scholarship from the Jacques Hadamard Mathematics Foundation (FMJH). Before my PhD, I worked across quantitative research and applied AI, with roles at BNP Paribas, SociĂ©tĂ© GĂ©nĂ©rale, and CrĂ©dit Agricole, and as a research scientist at JACOBB (Centre for Applied AI).
My primary research develops scalable low-rank variational inference methods (W ≈ ABT) for Bayesian neural networks, with provable guarantees and an emphasis on parameter efficiency. I also study whether these structures let networks automatically adapt their complexity to the data generating process (Occam's Razor).
A low-rank variational inference framework for Bayesian neural networks that parameterizes weights as W = ABT, reducing parameter complexity from O(mn) to O((m+n)r). This induces a posterior singular with respect to the Lebesgue measure, capturing structured weight correlations through shared latent factors. We derive PAC-Bayes generalization bounds and prove loss bounds via the Eckart-Young-Mirsky theorem. Empirically, the method matches 5-member Deep Ensembles with up to 15× fewer parameters and substantially improves out-of-distribution detection across MLPs, LSTMs, and Transformers.
We decompose mutual information into a per-class uncertainty vector Ck(x) = σk² / (2μk), letting safety-critical classifiers distinguish where a model is uncertain, not just how much. Applied to diabetic retinopathy grading, the critical-class score reduces selective risk by 34.7% over mutual information and 56.2% over variance baselines.
We study when a structured low-rank Gaussian variational posterior can certify a deterministic predictor in a Bayesian neural network with factorized layers Wi = AiBiT. The same model yields three natural certification targets, the posterior Gibbs predictor, the posterior predictive mean, and a deterministic center network; this paper focuses on the deterministic-center route.
A variational inference framework that factorizes weight matrices to reduce parameter complexity from O(n2) to O(n), enabling uncertainty quantification in large-scale models.
Applied deep neural networks to solve high-dimensional PDEs for Credit Valuation Adjustment, overcoming the curse of dimensionality faced by finite-difference methods.
Implemented statistical techniques to estimate the kernel of Volterra processes, validating the rough-volatility hypothesis in financial time series.
TA for MATH 139 (Calculus with Precalculus) and MATH 141 (Calculus 2): leading problem-solving tutorials and strengthening students' conceptual grasp of limits, differentiation, integration, and series.
Designed and executed rigorous evaluation protocols for quantitative and machine learning models across the group, probing robustness, calibration, and governance beyond standard validation. Assessed model sensitivity to input perturbations and distributional shift, and translated findings into remediation recommendations for senior management and group risk committees.
Applied R&D in NLP and Reinforcement Learning. Built a reinforcement learning agent for advertising budget pacing and a collaborative-filtering matchmaking system for startups, and deployed models from prototype to production.
Selected for a competitive mentorship program with in-depth exposure to quantitative research methodologies, financial modeling, and career pathways in global markets.
Contributed to the Haussmann Project: developed and calibrated Probability of Default and Loss Given Default models for low-default portfolios.
Built mathematical models quantifying the bank's exposure to Central Counterparty Clearing Houses, modeling default-fund contributions with Monte Carlo simulation.
I am an active member of Women in AI North America and the Association of Women in Mathematics. I mentor at Jiggen In STEM, supporting high school girls in Senegal, and served as the former representative of AFSA Québec at McGill.
