Vincent Quenneville-Bélair, phd

Machine Learning Scientist

Supply Chain Optimization Technologies, Amazon

2021 - Present

Build a autonomous search over optimization policies, using a Lean kernel to certify candidates and promote provably best ones. The final algorithm cuts placement cost 10% and runtime from >2 hours to <2 minutes.

Prototype predictive model for inventory placement across >100 warehouses, accelerating pipelines by 10x by reframing decision-making into a forecasting formulation.

Lead vision and production launch of multimodal foundation forecasting model across >100M products, achieving zero-shot accuracy gains of >100 bps for external partners.

Use attention mechanism to design distribution-free model to forecast counterfactual demand for supply chain optimization; lifted profit >6% with latency <20 ms.

Leverage LLM embeddings and approximate nearest-neighbor retrieval to capture cross-product demand effects, sharing demand signal across semantically similar products, outperforming by >50 bps over production.

Develop RLSF (Reinforcement Learning from Sales Feedback) to align LLM-generated product descriptions to sales outcomes, using A/B experiment data for reward modeling.

Machine Learning Engineer

PyTorch, Facebook AI

2019 - 2021

Create torchaudio library, engaging over 200 open source contributors.

Develop speech-to-text and text-to-speech deep learning models, implementing efficient C++ and CUDA autograd operators, such as differentiable RNN-Transducer loss.

Integrate novel optimizers collaborating with academic and industrial researchers.

Machine Learning Scientist

Supply Chain Forecasting and Marketing Measurement, Amazon

2017 - 2019

Develop long-term strategic capacity forecast to generate supply chain expansion recommendations.

Build causal impact models to estimate advertising lift and optimize advertising portfolio.

Teach courses on matrix factorization, dimensionality reduction, recommender systems.

Chief Data Scientist

Vizanda

2016 - 2017

Build real-time software to automatically extract information, visualizations, insights from unstructured data.

Chu Assistant Professor of Applied Mathematics

Columbia University

2015 - 2017

Develop numerical simulations for physical systems using novel finite elements. Model gravitational waves to understand black hole collisions.