Thierry Gross

Certification Deep Learning at École Polytechnique (obtained in Déc. 2024): – Advanced techniques in NLP, Computer vision, Recommendation Systems, Graphs, etc. – Transformers, LLM, Generative AI, Graph Neural Networks, CNN – Python, Pytorch, Document embeddings, Graph embeddings, Collaborative filtering, PEFT, etc. – Full program: https://info.exed.polytechnique.edu/Advanced-AI-for-Data-Analysis Mission at Lefebvre-Dalloz (2023): – Audit of the existing process of storage solution and provision of case law – Adaptation considering the expected increasing quantity of data – Performance of a full-database search tool using PostgreSQL

As a NLP Engineer in the Natural Language Understanding team, I contributed to improving Siri's understanding and accuracy in French by enhancing both language models and data workflows. Key contributions: – Supervision of annotation task to enhance training data – Dialog generation: pattern recognition, linguistic resources (dictionaries, grammars, etc.) – Synthetic data generation pipelines to enrich training data (template-based, inflection-based methods) – Monitored pipeline delivery to ensure reliable data and model integration – Recruitment of computational linguists – Responsible for the "Social conversation" part of SIRI – Machine Learning: Built classification and diagnostic tools to identify and resolve task-related bugs – Python, Javascript, Java, Bash, Gradle, Jenkins, LSTM, Multi-label Classification

As Project Manager, I was responsible for the technical coordination of European projects for institutional clients (European Commission, European GNSS Agency, European Space Agency, CNES) and industrial consortiums. I worked on topics such as Galileo, embedded development, automotive applications, sensor-fusion models, system security, technical analyses, certification, and standardisation activities in ITS applications. Main achievement was the Design and Development of the first secured GNSS receiver for ITS applications: – Product definition, technical and functional specifications – Definition of confidence in positioning – Definition of attacks and validation of the countermeasures – Design of security functions (electronic signature, authentication mechanisms, PKI, etc.) – Sensor (MEMS) fusion with GNSS: statistical modelling in R, real-time (C, C++) – Software architecture and implementation on ARM (Linux, OS20, C, C++) – Certification and standardization activities Tech stack: C, C++, Python, R, SQLite, Perl, ARM, PKI, OpenSSL, Teseo 3, ST33, etc.

Prima Solutions (formerly EFFISOFT) is a software publisher specialised in reinsurance workflows. Its main clients are international insurance and reinsurance companies. I was responsible for: – Technical and functional consulting for software evolution and maintenance – Custom developments for clients (AXA, VinaRe, etc.) – Oracle/PLSQL development and SQL performance tuning (architecture and code optimisation)

Lectures for undergraduate and graduate students (BTS Electronic Systems, Master Telecom, etc.). Responsible for the preparation for the competition examination at Telecom Paristech.

Within SGAM AI (Hedge fund of Société Générale Asset Management Alternative Investment): engineer in charge of Volatility Arbitrage desk to assess new investment strategies based on statistical analysis.

Research and Teacher at the Laboratory "Probabilités et modèles aléatoires" in Paris 6 and Paris 7. – Teaching: Statistics, Algebra and Analysis – Research: Random walk in random environment, product of random matrices, multiplicative chaos, random trees, Lyapunov exponents – Thesis: We investigate the asymptotic behavior of a particle performing a random walk on a graph within a random (inhomogeneous) environment. Such randomness introduces new effects not seen in uniform settings, notably a slowdown caused by trapping regions in the environment. While random walks in one-dimensional random environments are well understood, extending these results to more complex graphs raises major challenges. I focus on a model called "random strings," which offers an intermediate setting — more intricate than one-dimensional graphs but with a simpler form of randomness — enabling a formal connection with classical one-dimensional models.