Developed tools to extract new and clear-cut information related to ESG topic within financial and extra-financial corporate reports.

Why?
Offset the lack of quality data among ESG indicator to enhanceinsights and metrics of business-related topics such as portfolioalignment, physical risk assessment...
• Refined corporate NACE classification using siamestransformers.
• Created a fine-tuned entity extraction tool to precisely extractrelevant information related to assets, geolocations, metrics…within corporate reports.
• Leveraged few-shot learning techniques to detect sectorial
• Publication of a Climate Finance Benchmark, assessing the relevance of RAG approaches to extract corporate climate related information.

Implemented a credit risk score using banking data for clients with no credit history.

Why?
Serve more widely clients with no credit history and use the information included in banking data for credit worthiness and repayment behaviour.
• Analyzed the sociodemographic and risk profile of clients with no credit history.
• Implemented a pipeline to process banking data and extract relevant KPIs using domain knowledge and statistical characteristics.
• Developed a reusable methodology for integrating banking data in Machine Learning models.

Leveraged state-of-the-art graph neural network approaches to improve suspicious activity detection.

Why?
Tackle the increasing regulation pressure and continuous surveillance needs regarding anti-money laundering and counter terrorist financing.
• Benchmarked existing supervised, unsupervised and weakly supervised techniques to detected suspicious activities.
• Open-sourced ad-hoc study to assess the relevance of graph neural networks applied to incomplete transactional graphs.
• Improved existing detection tool by capitalizing on graph autoencoded embeddings with operational and conclusive results.

Developed predictive models of sinistrality for a car insurance portfolio. Highlighted the explaining factors.

Why?
Include them into the partner's prevention roadmap to reduce sinistrality.
• Deep-dived on data to understand and aggregate them at the appropriate level.
• Collaborated with academic experts on statistics for actuarial science to challenge benchmark actuarial models for sinistrality with machine learning.
• Shared an interactive support tool to understand sinistrality drivers.