Clément Olivier

At Miuros the datascience team leverages AI to enhance digital tools to help customer services through data driven approaches. My role in the team:

- Explore state-of-art NLP technologies, in particular unsupervised learning, with the aim of designing a new product meant to discover trending topics mentionned in customer messages. Closely working with the customer team.

- Test, implement and provide through API, NLP proof-of-concept tools to expand products.

- Implement interactive user interface to investigate the benefit of human-in the-loop processes.

- Improve in-production models (Classification, Recommender System) in terms of technology, accuracy, computational cost and monitoring.

- Onboard new customer: configure, train, evaluate and deploy models. NLP technologies:

-Text clustering (Umap + HDBscan, Auto-Encoders)

-Few shot learning (Prototypical Network)

-Language Model (Bert, Sentence transformer, ~FastText)

-Keyword Extraction

-Topic Modeling (Correx) Stack: Pytorch, Tensorflow, Transformer, Scikit-Learn, Spacy, FastAPI, VueJS, Docker, AWS and more

My work consists in improving and implementing new components of our Natural Language Understanding model. Some projects :

- Implementation of a ML algorithms to improve the relevance our email suggestion response module (Python)

- Improvement of our chatbot workflow to manage more complex scenarios (C ++) Keywords : ✔ Machine Learning ✔ Natural Language Understanding ✔ Automated Response Suggestion ✔ Chatbots ✔ Data Analysis ✔ Intention Detection

I proposed during my PhD thesis an innovative machine learning algorithm enabling to construct surrogate models for highly expensive physical models. The methodology is based on a tensor decomposition framework that I have developed. The method is applied to approximate outputs of interest of mechanical simulations involving non-linear constitutive material laws. I have implemented from scratch my predictive model in Python and designed an interactive and collaborative tool in Javascript to visualize in real-time the outputs and compare it with the physical data. This tool is used by experts in material science to efficiently understand and manipulate the outputs of physical models. Keywords : ✔ Machine learning ✔ Approximation of highly expensive model ✔ Development of numerical tools for experts ✔ Big data ✔ Tensor decomposition techniques ✔ Data-intensive computing During my PhD, I attended and presented my work at various international conferences: Projection Based Model Reduction, Oberwolfach, Germany Tensor Decompositions and Applications, Leuven, Belgium Workshop on Order Reduction Methods, Bad Herrenald, Germany U.S. National Congress on Computational Mechanics, Montréal, Canada