Arnault Gombert

The Development Impact Group’s Artificial Intelligence Team (DIME AI) is pioneering the next frontier of impact evaluation and development programming. By leveraging AI and machine learning, our applied AI lab develops and implements research, interventions, and tools that drive AI-powered solutions to leapfrog development gaps across key thematic areas. At the core of our work is a commitment to maintaining and benchmarking the highest ethical standards and rigor, particularly when deploying AI in vulnerable communities to address pressing global challenges. Guided by this principle, DIME AI leads several flagship initiatives, including: - ImpactAI: Synthesizing quantitative causal research evidence to improve development financing decisions. - ZeroHungerAI: Advancing predictions of food crises outbreaks with non-traditional data sources and state-of-the-art models. - SocialAI: Developing AI algorithms for low-resource contexts and languages, including detecting online hate speech and mitigating biases in large-language models (LLMs).

Advanced Methods in Natural Language Processing - Masters: Data Science for Decision Making and Data Science Methdology - 10 sessions of 2 hours for 30 international master students - Main topics: RNN, Attention, Transformers, BERT models

As a NLP freelancer I worked on different projects with international clients: • Created an event detection algorithm from texts • Audit deep learning model for a social services institution to detect bias • Wrote a methodology to audit RNN outputs based on interviews for a legal tech • Clustered protected groups from the interviews with BERT embeddings • Analyzed the biases of the models regarding protected groups • Created a NLP classifier to detect hatred text on social networks • Built a tool to eliminate symmetric biases in annotations tasks

A computational medicine company that uses latest data mining process and machine learning to extract and structure medical data contained in medical records, medical notes or scientific publications among others. I used to work as a Data Scientist mainly on operational research: • Developed a RNN algorithm to predict why a patient will come back in a hospital based on EHRs (Electronic Health Records - Medical History) with Tensorflow and Keras • Modeled when a patients in traumatology should come back to hospital with scikit-learn • Wrote an article about the RNN solutions we created at IOMED • Organized and automatized the dataset creation process with prodigy (from raw data to consolidated dataset for training/validating/testing) • Used dependency parsing algorithm with spaCy on medical data to extract relations between medical concepts • Added the dependency parsing algorithm to IOMED API to improve it • Developing a word generator with Tensorflow at character and word level with RNN for text generation • Created a dilated CNN algorithm for contextualized word embeddings with Keras

Data Scientist & Start-up manager @ZEBOX incubator.

Optimizing recycling mixing data and operational analysis

• Estimated assets flow models through time for a national bank. Those models will need to be monitored by the bank, nevertheless the bank can manage their gaps in the middle term. • Built scoring and ALM tools for a national bank. I used VBA, related to the models I first built. I did it for the ALM models and the scoring models. • Back-tested grant and behavior scoring models for a national bank. I used out of sample tests, or out of time tests. Those tests put forward the stability of the models (structural stability and time stability), and now even if the bank still has to monitor the models, it's efficient at least in the middle term. • Wrote a global methodology about the VaR calculations. The article is in the new GRA booklet. • Wrote a white paper on population migrations of a portfolio during a stressed period. I used epidemiological processes in order to fit the stress growth in the portfolio (the population).

• Audited risk models of a French systemic Bank. • Analyzed data of SME costumers for a national bank: retreatment, tree algorithms, out of sample tests. I had to clean the dataset the bank gave us in order to treat data. After we made descriptive statistics and used Chaid and C&RT algorithms to class people in each variable in function of default. • Created a scoring model, retail banking, and ranked risk for a national bank: out of sample tests, bootstrapping, out of time tests. After analyzing data, I had to choose ones i used to create the score and use logistic regression to class people for loans. And I had to make tests on my risk model. • Created a stress-testing model for a specialized financial institution. • Wrote a white paper on a methodology to grade post-stress PD on any kind of portfolios.

Estimated customers’ incomes and risks of repayment. Discussed the alternative for late delay payers: I met clients almost each day to understand why there is a delay, and how customers should process for repayment. Made basic audit and intern controls of officers’ work: I controlled the numbers officers declared when they made "tontines". Described customers’ life and activities for WAGES financial partner sitcom: Everybody can lend money to WAGES customers without interest when you use kiva site web. I promoted activities of WAGES clients to get those loans.

First Experience in an international Bank. Data management and filled files in order to rate over-indebtedness customers.