- Designed, implemented and optimized the training methodology for ultra-low power Keyword Spotting and Natural Language Understanding models, achieving high recognition rate in very noisy and far-field conditions. - Conducted research study on SOTA compression techniques (sparsity, quantization) for language models- Contributed to the standardization of ML software best practices across the organization - Sourced data providers, negotiated prices, established scope and specifications of clients' projects.- Designed a scalable computation scheme to distribute Attention layers on a Network-on-Chip.

- Built a production-grade AI-based Speech Enhancement algorithm (preserving speech while removing noise), demonstrated at CES 2022, 2023 and 2024, targeted at tiny embedded devices.- Designed novel architectures to achieve ultra-low latency and power for Speech Enhancement.- Conducted research on innovative loss function design to preserve speech naturalness.- Addressed and resolved various challenges in embedded deep learning for audio, including reverberation, far-field effects, quantization, and artifacts from spectral processing.- Established processes and metrics for model performance evaluation.- Conducted research on quantization, weight and activation sparsity and its impact on the performance vs power/latency trade-off on TinyML benchmarks.- Developed hardware-optimized kernels for convolution layers.- One of the main contributors to the fmot model optimization library, including: hardware-aware pruning techniques, quantization, quantization debugging tools...

- Developed a statistical methodology to compare strategy performances on auto-correlated time series- Built ML models and evaluation frameworks for factor-based alpha strategies (equity)- Designed features and models for earnings forecasting

Taught CME100, Vector Calculus for Engineers (including optimization, multi-dimension integrals, linear algebra…) in Q3 2019 and 2020.

Part of Quantitative Investment Strategies, whose aim is to propose new quantitative systematic strategies (equity and cross-asset) to institutional clients:- Worked on SG's flagship hedging strategies, implemented a put replication backtester- Developed a new short vol strategy based on Var Swaps- Optimization of Python code runtime- Developed a VIX-based strategy (worked on VIX modelization, proposed new ratios and systematic methodologies to account for hedging performances, challenged and optimized input parameters)- Reviewed SG's Cross-Asset Trend Following Strategy, proposed new signal detections and developed a new weight allocation based on the risk parity approach

Part of the Global Financial Resources Management Team, in charge of optimizing regulatory margins impacting trading books:- Proposed several clearing and compression packages, conducted analysis (RWA reduction, IM optimization, CSA impact)- Priced floors in the context of negative rates issues (EONIA/Euribor cost on collateral for bilateral positions)- Developed a pricing framework for sales people (raw IM & RWA costs/savings and CSA impact for minor trades)- Developed reporting tools on HSBC's credit and rate products- Reviewed new Initial Margin computation methodologies in clearing houses for rate products

Providing students in courses preparing for top French Business Schools with weekly oral exams, called "kholles", in analysis, probabilities and algebra